2 * @brief Lightning memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2016 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
38 #if defined(__WIN64__)
39 #define _FILE_OFFSET_BITS 64
44 #include <wchar.h> /* get wcscpy() */
46 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
47 * as int64 which is wrong. MSVC doesn't define it at all, so just
51 #define MDB_THR_T DWORD
52 #include <sys/types.h>
55 # include <sys/param.h>
57 # define LITTLE_ENDIAN 1234
58 # define BIG_ENDIAN 4321
59 # define BYTE_ORDER LITTLE_ENDIAN
61 # define SSIZE_MAX INT_MAX
65 #include <sys/types.h>
67 #define MDB_PID_T pid_t
68 #define MDB_THR_T pthread_t
69 #include <sys/param.h>
72 #ifdef HAVE_SYS_FILE_H
78 #if defined(__mips) && defined(__linux)
79 /* MIPS has cache coherency issues, requires explicit cache control */
80 #include <asm/cachectl.h>
81 extern int cacheflush(char *addr, int nbytes, int cache);
82 #define CACHEFLUSH(addr, bytes, cache) cacheflush(addr, bytes, cache)
84 #define CACHEFLUSH(addr, bytes, cache)
87 #if defined(__linux) && !defined(MDB_FDATASYNC_WORKS)
88 /** fdatasync is broken on ext3/ext4fs on older kernels, see
89 * description in #mdb_env_open2 comments. You can safely
90 * define MDB_FDATASYNC_WORKS if this code will only be run
91 * on kernels 3.6 and newer.
93 #define BROKEN_FDATASYNC
107 typedef SSIZE_T ssize_t;
112 #if defined(__sun) || defined(ANDROID)
113 /* Most platforms have posix_memalign, older may only have memalign */
114 #define HAVE_MEMALIGN 1
118 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
119 #include <netinet/in.h>
120 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
123 #if defined(__APPLE__) || defined (BSD)
124 # define MDB_USE_POSIX_SEM 1
125 # define MDB_FDATASYNC fsync
126 #elif defined(ANDROID)
127 # define MDB_FDATASYNC fsync
133 #ifdef MDB_USE_POSIX_SEM
134 # define MDB_USE_HASH 1
135 #include <semaphore.h>
137 #define MDB_USE_POSIX_MUTEX 1
141 #if defined(_WIN32) + defined(MDB_USE_POSIX_SEM) \
142 + defined(MDB_USE_POSIX_MUTEX) != 1
143 # error "Ambiguous shared-lock implementation"
147 #include <valgrind/memcheck.h>
148 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
149 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
150 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
151 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
152 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
154 #define VGMEMP_CREATE(h,r,z)
155 #define VGMEMP_ALLOC(h,a,s)
156 #define VGMEMP_FREE(h,a)
157 #define VGMEMP_DESTROY(h)
158 #define VGMEMP_DEFINED(a,s)
162 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
163 /* Solaris just defines one or the other */
164 # define LITTLE_ENDIAN 1234
165 # define BIG_ENDIAN 4321
166 # ifdef _LITTLE_ENDIAN
167 # define BYTE_ORDER LITTLE_ENDIAN
169 # define BYTE_ORDER BIG_ENDIAN
172 # define BYTE_ORDER __BYTE_ORDER
176 #ifndef LITTLE_ENDIAN
177 #define LITTLE_ENDIAN __LITTLE_ENDIAN
180 #define BIG_ENDIAN __BIG_ENDIAN
183 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
184 #define MISALIGNED_OK 1
190 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
191 # error "Unknown or unsupported endianness (BYTE_ORDER)"
192 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
193 # error "Two's complement, reasonably sized integer types, please"
197 /** Put infrequently used env functions in separate section */
199 # define ESECT __attribute__ ((section("__TEXT,text_env")))
201 # define ESECT __attribute__ ((section("text_env")))
208 #define CALL_CONV WINAPI
213 /** @defgroup internal LMDB Internals
216 /** @defgroup compat Compatibility Macros
217 * A bunch of macros to minimize the amount of platform-specific ifdefs
218 * needed throughout the rest of the code. When the features this library
219 * needs are similar enough to POSIX to be hidden in a one-or-two line
220 * replacement, this macro approach is used.
224 /** Features under development */
229 /** Wrapper around __func__, which is a C99 feature */
230 #if __STDC_VERSION__ >= 199901L
231 # define mdb_func_ __func__
232 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
233 # define mdb_func_ __FUNCTION__
235 /* If a debug message says <mdb_unknown>(), update the #if statements above */
236 # define mdb_func_ "<mdb_unknown>"
239 /* Internal error codes, not exposed outside liblmdb */
240 #define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10)
242 #define MDB_OWNERDEAD ((int) WAIT_ABANDONED)
243 #elif defined(MDB_USE_POSIX_MUTEX) && defined(EOWNERDEAD)
244 #define MDB_OWNERDEAD EOWNERDEAD /**< #LOCK_MUTEX0() result if dead owner */
248 #define GLIBC_VER ((__GLIBC__ << 16 )| __GLIBC_MINOR__)
250 /** Some platforms define the EOWNERDEAD error code
251 * even though they don't support Robust Mutexes.
252 * Compile with -DMDB_USE_ROBUST=0, or use some other
253 * mechanism like -DMDB_USE_POSIX_SEM instead of
254 * -DMDB_USE_POSIX_MUTEX.
255 * (Posix semaphores are not robust.)
257 #ifndef MDB_USE_ROBUST
258 /* Android currently lacks Robust Mutex support. So does glibc < 2.4. */
259 # if defined(MDB_USE_POSIX_MUTEX) && (defined(ANDROID) || \
260 (defined(__GLIBC__) && GLIBC_VER < 0x020004))
261 # define MDB_USE_ROBUST 0
263 # define MDB_USE_ROBUST 1
265 #endif /* !MDB_USE_ROBUST */
267 #if defined(MDB_USE_POSIX_MUTEX) && (MDB_USE_ROBUST)
268 /* glibc < 2.12 only provided _np API */
269 # if (defined(__GLIBC__) && GLIBC_VER < 0x02000c) || \
270 (defined(PTHREAD_MUTEX_ROBUST_NP) && !defined(PTHREAD_MUTEX_ROBUST))
271 # define PTHREAD_MUTEX_ROBUST PTHREAD_MUTEX_ROBUST_NP
272 # define pthread_mutexattr_setrobust(attr, flag) pthread_mutexattr_setrobust_np(attr, flag)
273 # define pthread_mutex_consistent(mutex) pthread_mutex_consistent_np(mutex)
275 #endif /* MDB_USE_POSIX_MUTEX && MDB_USE_ROBUST */
277 #if defined(MDB_OWNERDEAD) && (MDB_USE_ROBUST)
278 #define MDB_ROBUST_SUPPORTED 1
282 #define MDB_USE_HASH 1
283 #define MDB_PIDLOCK 0
284 #define THREAD_RET DWORD
285 #define pthread_t HANDLE
286 #define pthread_mutex_t HANDLE
287 #define pthread_cond_t HANDLE
288 typedef HANDLE mdb_mutex_t, mdb_mutexref_t;
289 #define pthread_key_t DWORD
290 #define pthread_self() GetCurrentThreadId()
291 #define pthread_key_create(x,y) \
292 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
293 #define pthread_key_delete(x) TlsFree(x)
294 #define pthread_getspecific(x) TlsGetValue(x)
295 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
296 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
297 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
298 #define pthread_cond_signal(x) SetEvent(*x)
299 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
300 #define THREAD_CREATE(thr,start,arg) \
301 (((thr) = CreateThread(NULL, 0, start, arg, 0, NULL)) ? 0 : ErrCode())
302 #define THREAD_FINISH(thr) \
303 (WaitForSingleObject(thr, INFINITE) ? ErrCode() : 0)
304 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
305 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
306 #define mdb_mutex_consistent(mutex) 0
307 #define getpid() GetCurrentProcessId()
308 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
309 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
310 #define ErrCode() GetLastError()
311 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
312 #define close(fd) (CloseHandle(fd) ? 0 : -1)
313 #define munmap(ptr,len) UnmapViewOfFile(ptr)
314 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
315 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
317 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
321 #define THREAD_RET void *
322 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
323 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
324 #define Z "z" /**< printf format modifier for size_t */
326 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
327 #define MDB_PIDLOCK 1
329 #ifdef MDB_USE_POSIX_SEM
331 typedef sem_t *mdb_mutex_t, *mdb_mutexref_t;
332 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
333 #define UNLOCK_MUTEX(mutex) sem_post(mutex)
336 mdb_sem_wait(sem_t *sem)
339 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
343 #else /* MDB_USE_POSIX_MUTEX: */
344 /** Shared mutex/semaphore as the original is stored.
346 * Not for copies. Instead it can be assigned to an #mdb_mutexref_t.
347 * When mdb_mutexref_t is a pointer and mdb_mutex_t is not, then it
348 * is array[size 1] so it can be assigned to the pointer.
350 typedef pthread_mutex_t mdb_mutex_t[1];
351 /** Reference to an #mdb_mutex_t */
352 typedef pthread_mutex_t *mdb_mutexref_t;
353 /** Lock the reader or writer mutex.
354 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
356 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
357 /** Unlock the reader or writer mutex.
359 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
360 /** Mark mutex-protected data as repaired, after death of previous owner.
362 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex)
363 #endif /* MDB_USE_POSIX_SEM */
365 /** Get the error code for the last failed system function.
367 #define ErrCode() errno
369 /** An abstraction for a file handle.
370 * On POSIX systems file handles are small integers. On Windows
371 * they're opaque pointers.
375 /** A value for an invalid file handle.
376 * Mainly used to initialize file variables and signify that they are
379 #define INVALID_HANDLE_VALUE (-1)
381 /** Get the size of a memory page for the system.
382 * This is the basic size that the platform's memory manager uses, and is
383 * fundamental to the use of memory-mapped files.
385 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
388 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
391 #define MNAME_LEN (sizeof(pthread_mutex_t))
396 #ifdef MDB_ROBUST_SUPPORTED
397 /** Lock mutex, handle any error, set rc = result.
398 * Return 0 on success, nonzero (not rc) on error.
400 #define LOCK_MUTEX(rc, env, mutex) \
401 (((rc) = LOCK_MUTEX0(mutex)) && \
402 ((rc) = mdb_mutex_failed(env, mutex, rc)))
403 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc);
405 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
406 #define mdb_mutex_failed(env, mutex, rc) (rc)
410 /** A flag for opening a file and requesting synchronous data writes.
411 * This is only used when writing a meta page. It's not strictly needed;
412 * we could just do a normal write and then immediately perform a flush.
413 * But if this flag is available it saves us an extra system call.
415 * @note If O_DSYNC is undefined but exists in /usr/include,
416 * preferably set some compiler flag to get the definition.
420 # define MDB_DSYNC O_DSYNC
422 # define MDB_DSYNC O_SYNC
427 /** Function for flushing the data of a file. Define this to fsync
428 * if fdatasync() is not supported.
430 #ifndef MDB_FDATASYNC
431 # define MDB_FDATASYNC fdatasync
435 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
446 /** A page number in the database.
447 * Note that 64 bit page numbers are overkill, since pages themselves
448 * already represent 12-13 bits of addressable memory, and the OS will
449 * always limit applications to a maximum of 63 bits of address space.
451 * @note In the #MDB_node structure, we only store 48 bits of this value,
452 * which thus limits us to only 60 bits of addressable data.
454 typedef MDB_ID pgno_t;
456 /** A transaction ID.
457 * See struct MDB_txn.mt_txnid for details.
459 typedef MDB_ID txnid_t;
461 /** @defgroup debug Debug Macros
465 /** Enable debug output. Needs variable argument macros (a C99 feature).
466 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
467 * read from and written to the database (used for free space management).
473 static int mdb_debug;
474 static txnid_t mdb_debug_start;
476 /** Print a debug message with printf formatting.
477 * Requires double parenthesis around 2 or more args.
479 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
480 # define DPRINTF0(fmt, ...) \
481 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
483 # define DPRINTF(args) ((void) 0)
485 /** Print a debug string.
486 * The string is printed literally, with no format processing.
488 #define DPUTS(arg) DPRINTF(("%s", arg))
489 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
491 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
494 /** @brief The maximum size of a database page.
496 * It is 32k or 64k, since value-PAGEBASE must fit in
497 * #MDB_page.%mp_upper.
499 * LMDB will use database pages < OS pages if needed.
500 * That causes more I/O in write transactions: The OS must
501 * know (read) the whole page before writing a partial page.
503 * Note that we don't currently support Huge pages. On Linux,
504 * regular data files cannot use Huge pages, and in general
505 * Huge pages aren't actually pageable. We rely on the OS
506 * demand-pager to read our data and page it out when memory
507 * pressure from other processes is high. So until OSs have
508 * actual paging support for Huge pages, they're not viable.
510 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
512 /** The minimum number of keys required in a database page.
513 * Setting this to a larger value will place a smaller bound on the
514 * maximum size of a data item. Data items larger than this size will
515 * be pushed into overflow pages instead of being stored directly in
516 * the B-tree node. This value used to default to 4. With a page size
517 * of 4096 bytes that meant that any item larger than 1024 bytes would
518 * go into an overflow page. That also meant that on average 2-3KB of
519 * each overflow page was wasted space. The value cannot be lower than
520 * 2 because then there would no longer be a tree structure. With this
521 * value, items larger than 2KB will go into overflow pages, and on
522 * average only 1KB will be wasted.
524 #define MDB_MINKEYS 2
526 /** A stamp that identifies a file as an LMDB file.
527 * There's nothing special about this value other than that it is easily
528 * recognizable, and it will reflect any byte order mismatches.
530 #define MDB_MAGIC 0xBEEFC0DE
532 /** The version number for a database's datafile format. */
533 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
534 /** The version number for a database's lockfile format. */
535 #define MDB_LOCK_VERSION 1
537 /** @brief The max size of a key we can write, or 0 for computed max.
539 * This macro should normally be left alone or set to 0.
540 * Note that a database with big keys or dupsort data cannot be
541 * reliably modified by a liblmdb which uses a smaller max.
542 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
544 * Other values are allowed, for backwards compat. However:
545 * A value bigger than the computed max can break if you do not
546 * know what you are doing, and liblmdb <= 0.9.10 can break when
547 * modifying a DB with keys/dupsort data bigger than its max.
549 * Data items in an #MDB_DUPSORT database are also limited to
550 * this size, since they're actually keys of a sub-DB. Keys and
551 * #MDB_DUPSORT data items must fit on a node in a regular page.
553 #ifndef MDB_MAXKEYSIZE
554 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
557 /** The maximum size of a key we can write to the environment. */
559 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
561 #define ENV_MAXKEY(env) ((env)->me_maxkey)
564 /** @brief The maximum size of a data item.
566 * We only store a 32 bit value for node sizes.
568 #define MAXDATASIZE 0xffffffffUL
571 /** Key size which fits in a #DKBUF.
574 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
577 * This is used for printing a hex dump of a key's contents.
579 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
580 /** Display a key in hex.
582 * Invoke a function to display a key in hex.
584 #define DKEY(x) mdb_dkey(x, kbuf)
590 /** An invalid page number.
591 * Mainly used to denote an empty tree.
593 #define P_INVALID (~(pgno_t)0)
595 /** Test if the flags \b f are set in a flag word \b w. */
596 #define F_ISSET(w, f) (((w) & (f)) == (f))
598 /** Round \b n up to an even number. */
599 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
601 /** Used for offsets within a single page.
602 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
605 typedef uint16_t indx_t;
607 /** Default size of memory map.
608 * This is certainly too small for any actual applications. Apps should always set
609 * the size explicitly using #mdb_env_set_mapsize().
611 #define DEFAULT_MAPSIZE 1048576
613 /** @defgroup readers Reader Lock Table
614 * Readers don't acquire any locks for their data access. Instead, they
615 * simply record their transaction ID in the reader table. The reader
616 * mutex is needed just to find an empty slot in the reader table. The
617 * slot's address is saved in thread-specific data so that subsequent read
618 * transactions started by the same thread need no further locking to proceed.
620 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
622 * No reader table is used if the database is on a read-only filesystem, or
623 * if #MDB_NOLOCK is set.
625 * Since the database uses multi-version concurrency control, readers don't
626 * actually need any locking. This table is used to keep track of which
627 * readers are using data from which old transactions, so that we'll know
628 * when a particular old transaction is no longer in use. Old transactions
629 * that have discarded any data pages can then have those pages reclaimed
630 * for use by a later write transaction.
632 * The lock table is constructed such that reader slots are aligned with the
633 * processor's cache line size. Any slot is only ever used by one thread.
634 * This alignment guarantees that there will be no contention or cache
635 * thrashing as threads update their own slot info, and also eliminates
636 * any need for locking when accessing a slot.
638 * A writer thread will scan every slot in the table to determine the oldest
639 * outstanding reader transaction. Any freed pages older than this will be
640 * reclaimed by the writer. The writer doesn't use any locks when scanning
641 * this table. This means that there's no guarantee that the writer will
642 * see the most up-to-date reader info, but that's not required for correct
643 * operation - all we need is to know the upper bound on the oldest reader,
644 * we don't care at all about the newest reader. So the only consequence of
645 * reading stale information here is that old pages might hang around a
646 * while longer before being reclaimed. That's actually good anyway, because
647 * the longer we delay reclaiming old pages, the more likely it is that a
648 * string of contiguous pages can be found after coalescing old pages from
649 * many old transactions together.
652 /** Number of slots in the reader table.
653 * This value was chosen somewhat arbitrarily. 126 readers plus a
654 * couple mutexes fit exactly into 8KB on my development machine.
655 * Applications should set the table size using #mdb_env_set_maxreaders().
657 #define DEFAULT_READERS 126
659 /** The size of a CPU cache line in bytes. We want our lock structures
660 * aligned to this size to avoid false cache line sharing in the
662 * This value works for most CPUs. For Itanium this should be 128.
668 /** The information we store in a single slot of the reader table.
669 * In addition to a transaction ID, we also record the process and
670 * thread ID that owns a slot, so that we can detect stale information,
671 * e.g. threads or processes that went away without cleaning up.
672 * @note We currently don't check for stale records. We simply re-init
673 * the table when we know that we're the only process opening the
676 typedef struct MDB_rxbody {
677 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
678 * Multiple readers that start at the same time will probably have the
679 * same ID here. Again, it's not important to exclude them from
680 * anything; all we need to know is which version of the DB they
681 * started from so we can avoid overwriting any data used in that
682 * particular version.
684 volatile txnid_t mrb_txnid;
685 /** The process ID of the process owning this reader txn. */
686 volatile MDB_PID_T mrb_pid;
687 /** The thread ID of the thread owning this txn. */
688 volatile MDB_THR_T mrb_tid;
691 /** The actual reader record, with cacheline padding. */
692 typedef struct MDB_reader {
695 /** shorthand for mrb_txnid */
696 #define mr_txnid mru.mrx.mrb_txnid
697 #define mr_pid mru.mrx.mrb_pid
698 #define mr_tid mru.mrx.mrb_tid
699 /** cache line alignment */
700 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
704 /** The header for the reader table.
705 * The table resides in a memory-mapped file. (This is a different file
706 * than is used for the main database.)
708 * For POSIX the actual mutexes reside in the shared memory of this
709 * mapped file. On Windows, mutexes are named objects allocated by the
710 * kernel; we store the mutex names in this mapped file so that other
711 * processes can grab them. This same approach is also used on
712 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
713 * process-shared POSIX mutexes. For these cases where a named object
714 * is used, the object name is derived from a 64 bit FNV hash of the
715 * environment pathname. As such, naming collisions are extremely
716 * unlikely. If a collision occurs, the results are unpredictable.
718 typedef struct MDB_txbody {
719 /** Stamp identifying this as an LMDB file. It must be set
722 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
724 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
725 char mtb_rmname[MNAME_LEN];
727 /** Mutex protecting access to this table.
728 * This is the reader table lock used with LOCK_MUTEX().
730 mdb_mutex_t mtb_rmutex;
732 /** The ID of the last transaction committed to the database.
733 * This is recorded here only for convenience; the value can always
734 * be determined by reading the main database meta pages.
736 volatile txnid_t mtb_txnid;
737 /** The number of slots that have been used in the reader table.
738 * This always records the maximum count, it is not decremented
739 * when readers release their slots.
741 volatile unsigned mtb_numreaders;
744 /** The actual reader table definition. */
745 typedef struct MDB_txninfo {
748 #define mti_magic mt1.mtb.mtb_magic
749 #define mti_format mt1.mtb.mtb_format
750 #define mti_rmutex mt1.mtb.mtb_rmutex
751 #define mti_rmname mt1.mtb.mtb_rmname
752 #define mti_txnid mt1.mtb.mtb_txnid
753 #define mti_numreaders mt1.mtb.mtb_numreaders
754 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
757 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
758 char mt2_wmname[MNAME_LEN];
759 #define mti_wmname mt2.mt2_wmname
761 mdb_mutex_t mt2_wmutex;
762 #define mti_wmutex mt2.mt2_wmutex
764 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
766 MDB_reader mti_readers[1];
769 /** Lockfile format signature: version, features and field layout */
770 #define MDB_LOCK_FORMAT \
772 ((MDB_LOCK_VERSION) \
773 /* Flags which describe functionality */ \
774 + (((MDB_PIDLOCK) != 0) << 16)))
777 /** Common header for all page types. The page type depends on #mp_flags.
779 * #P_BRANCH and #P_LEAF pages have unsorted '#MDB_node's at the end, with
780 * sorted #mp_ptrs[] entries referring to them. Exception: #P_LEAF2 pages
781 * omit mp_ptrs and pack sorted #MDB_DUPFIXED values after the page header.
783 * #P_OVERFLOW records occupy one or more contiguous pages where only the
784 * first has a page header. They hold the real data of #F_BIGDATA nodes.
786 * #P_SUBP sub-pages are small leaf "pages" with duplicate data.
787 * A node with flag #F_DUPDATA but not #F_SUBDATA contains a sub-page.
788 * (Duplicate data can also go in sub-databases, which use normal pages.)
790 * #P_META pages contain #MDB_meta, the start point of an LMDB snapshot.
792 * Each non-metapage up to #MDB_meta.%mm_last_pg is reachable exactly once
793 * in the snapshot: Either used by a database or listed in a freeDB record.
795 typedef struct MDB_page {
796 #define mp_pgno mp_p.p_pgno
797 #define mp_next mp_p.p_next
799 pgno_t p_pgno; /**< page number */
800 struct MDB_page *p_next; /**< for in-memory list of freed pages */
802 uint16_t mp_pad; /**< key size if this is a LEAF2 page */
803 /** @defgroup mdb_page Page Flags
805 * Flags for the page headers.
808 #define P_BRANCH 0x01 /**< branch page */
809 #define P_LEAF 0x02 /**< leaf page */
810 #define P_OVERFLOW 0x04 /**< overflow page */
811 #define P_META 0x08 /**< meta page */
812 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
813 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
814 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
815 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
816 #define P_KEEP 0x8000 /**< leave this page alone during spill */
818 uint16_t mp_flags; /**< @ref mdb_page */
819 #define mp_lower mp_pb.pb.pb_lower
820 #define mp_upper mp_pb.pb.pb_upper
821 #define mp_pages mp_pb.pb_pages
824 indx_t pb_lower; /**< lower bound of free space */
825 indx_t pb_upper; /**< upper bound of free space */
827 uint32_t pb_pages; /**< number of overflow pages */
829 indx_t mp_ptrs[1]; /**< dynamic size */
832 /** Size of the page header, excluding dynamic data at the end */
833 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
835 /** Address of first usable data byte in a page, after the header */
836 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
838 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
839 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
841 /** Number of nodes on a page */
842 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
844 /** The amount of space remaining in the page */
845 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
847 /** The percentage of space used in the page, in tenths of a percent. */
848 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
849 ((env)->me_psize - PAGEHDRSZ))
850 /** The minimum page fill factor, in tenths of a percent.
851 * Pages emptier than this are candidates for merging.
853 #define FILL_THRESHOLD 250
855 /** Test if a page is a leaf page */
856 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
857 /** Test if a page is a LEAF2 page */
858 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
859 /** Test if a page is a branch page */
860 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
861 /** Test if a page is an overflow page */
862 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
863 /** Test if a page is a sub page */
864 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
866 /** The number of overflow pages needed to store the given size. */
867 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
869 /** Link in #MDB_txn.%mt_loose_pgs list.
870 * Kept outside the page header, which is needed when reusing the page.
872 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
874 /** Header for a single key/data pair within a page.
875 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
876 * We guarantee 2-byte alignment for 'MDB_node's.
878 * #mn_lo and #mn_hi are used for data size on leaf nodes, and for child
879 * pgno on branch nodes. On 64 bit platforms, #mn_flags is also used
880 * for pgno. (Branch nodes have no flags). Lo and hi are in host byte
881 * order in case some accesses can be optimized to 32-bit word access.
883 typedef struct MDB_node {
884 /** part of data size or pgno
886 #if BYTE_ORDER == LITTLE_ENDIAN
887 unsigned short mn_lo, mn_hi;
889 unsigned short mn_hi, mn_lo;
892 /** @defgroup mdb_node Node Flags
894 * Flags for node headers.
897 #define F_BIGDATA 0x01 /**< data put on overflow page */
898 #define F_SUBDATA 0x02 /**< data is a sub-database */
899 #define F_DUPDATA 0x04 /**< data has duplicates */
901 /** valid flags for #mdb_node_add() */
902 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
905 unsigned short mn_flags; /**< @ref mdb_node */
906 unsigned short mn_ksize; /**< key size */
907 char mn_data[1]; /**< key and data are appended here */
910 /** Size of the node header, excluding dynamic data at the end */
911 #define NODESIZE offsetof(MDB_node, mn_data)
913 /** Bit position of top word in page number, for shifting mn_flags */
914 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
916 /** Size of a node in a branch page with a given key.
917 * This is just the node header plus the key, there is no data.
919 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
921 /** Size of a node in a leaf page with a given key and data.
922 * This is node header plus key plus data size.
924 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
926 /** Address of node \b i in page \b p */
927 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
929 /** Address of the key for the node */
930 #define NODEKEY(node) (void *)((node)->mn_data)
932 /** Address of the data for a node */
933 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
935 /** Get the page number pointed to by a branch node */
936 #define NODEPGNO(node) \
937 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
938 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
939 /** Set the page number in a branch node */
940 #define SETPGNO(node,pgno) do { \
941 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
942 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
944 /** Get the size of the data in a leaf node */
945 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
946 /** Set the size of the data for a leaf node */
947 #define SETDSZ(node,size) do { \
948 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
949 /** The size of a key in a node */
950 #define NODEKSZ(node) ((node)->mn_ksize)
952 /** Copy a page number from src to dst */
954 #define COPY_PGNO(dst,src) dst = src
956 #if SIZE_MAX > 4294967295UL
957 #define COPY_PGNO(dst,src) do { \
958 unsigned short *s, *d; \
959 s = (unsigned short *)&(src); \
960 d = (unsigned short *)&(dst); \
967 #define COPY_PGNO(dst,src) do { \
968 unsigned short *s, *d; \
969 s = (unsigned short *)&(src); \
970 d = (unsigned short *)&(dst); \
976 /** The address of a key in a LEAF2 page.
977 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
978 * There are no node headers, keys are stored contiguously.
980 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
982 /** Set the \b node's key into \b keyptr, if requested. */
983 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
984 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
986 /** Set the \b node's key into \b key. */
987 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
989 /** Information about a single database in the environment. */
990 typedef struct MDB_db {
991 uint32_t md_pad; /**< also ksize for LEAF2 pages */
992 uint16_t md_flags; /**< @ref mdb_dbi_open */
993 uint16_t md_depth; /**< depth of this tree */
994 pgno_t md_branch_pages; /**< number of internal pages */
995 pgno_t md_leaf_pages; /**< number of leaf pages */
996 pgno_t md_overflow_pages; /**< number of overflow pages */
997 size_t md_entries; /**< number of data items */
998 pgno_t md_root; /**< the root page of this tree */
1001 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
1002 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
1003 /** #mdb_dbi_open() flags */
1004 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
1005 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
1007 /** Handle for the DB used to track free pages. */
1009 /** Handle for the default DB. */
1011 /** Number of DBs in metapage (free and main) - also hardcoded elsewhere */
1014 /** Number of meta pages - also hardcoded elsewhere */
1017 /** Meta page content.
1018 * A meta page is the start point for accessing a database snapshot.
1019 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
1021 typedef struct MDB_meta {
1022 /** Stamp identifying this as an LMDB file. It must be set
1025 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
1026 uint32_t mm_version;
1027 void *mm_address; /**< address for fixed mapping */
1028 size_t mm_mapsize; /**< size of mmap region */
1029 MDB_db mm_dbs[CORE_DBS]; /**< first is free space, 2nd is main db */
1030 /** The size of pages used in this DB */
1031 #define mm_psize mm_dbs[FREE_DBI].md_pad
1032 /** Any persistent environment flags. @ref mdb_env */
1033 #define mm_flags mm_dbs[FREE_DBI].md_flags
1034 /** Last used page in the datafile.
1035 * Actually the file may be shorter if the freeDB lists the final pages.
1038 volatile txnid_t mm_txnid; /**< txnid that committed this page */
1041 /** Buffer for a stack-allocated meta page.
1042 * The members define size and alignment, and silence type
1043 * aliasing warnings. They are not used directly; that could
1044 * mean incorrectly using several union members in parallel.
1046 typedef union MDB_metabuf {
1049 char mm_pad[PAGEHDRSZ];
1054 /** Auxiliary DB info.
1055 * The information here is mostly static/read-only. There is
1056 * only a single copy of this record in the environment.
1058 typedef struct MDB_dbx {
1059 MDB_val md_name; /**< name of the database */
1060 MDB_cmp_func *md_cmp; /**< function for comparing keys */
1061 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
1062 MDB_rel_func *md_rel; /**< user relocate function */
1063 void *md_relctx; /**< user-provided context for md_rel */
1066 /** A database transaction.
1067 * Every operation requires a transaction handle.
1070 MDB_txn *mt_parent; /**< parent of a nested txn */
1071 /** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */
1073 pgno_t mt_next_pgno; /**< next unallocated page */
1074 /** The ID of this transaction. IDs are integers incrementing from 1.
1075 * Only committed write transactions increment the ID. If a transaction
1076 * aborts, the ID may be re-used by the next writer.
1079 MDB_env *mt_env; /**< the DB environment */
1080 /** The list of pages that became unused during this transaction.
1082 MDB_IDL mt_free_pgs;
1083 /** The list of loose pages that became unused and may be reused
1084 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1086 MDB_page *mt_loose_pgs;
1087 /** Number of loose pages (#mt_loose_pgs) */
1089 /** The sorted list of dirty pages we temporarily wrote to disk
1090 * because the dirty list was full. page numbers in here are
1091 * shifted left by 1, deleted slots have the LSB set.
1093 MDB_IDL mt_spill_pgs;
1095 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1096 MDB_ID2L dirty_list;
1097 /** For read txns: This thread/txn's reader table slot, or NULL. */
1100 /** Array of records for each DB known in the environment. */
1102 /** Array of MDB_db records for each known DB */
1104 /** Array of sequence numbers for each DB handle */
1105 unsigned int *mt_dbiseqs;
1106 /** @defgroup mt_dbflag Transaction DB Flags
1110 #define DB_DIRTY 0x01 /**< DB was written in this txn */
1111 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1112 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1113 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1114 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */
1115 #define DB_DUPDATA 0x20 /**< DB is #MDB_DUPSORT data */
1117 /** In write txns, array of cursors for each DB */
1118 MDB_cursor **mt_cursors;
1119 /** Array of flags for each DB */
1120 unsigned char *mt_dbflags;
1121 /** Number of DB records in use, or 0 when the txn is finished.
1122 * This number only ever increments until the txn finishes; we
1123 * don't decrement it when individual DB handles are closed.
1127 /** @defgroup mdb_txn Transaction Flags
1131 /** #mdb_txn_begin() flags */
1132 #define MDB_TXN_BEGIN_FLAGS MDB_RDONLY
1133 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1134 /* internal txn flags */
1135 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1136 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */
1137 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1138 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1139 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1140 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */
1141 /** most operations on the txn are currently illegal */
1142 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD)
1144 unsigned int mt_flags; /**< @ref mdb_txn */
1145 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1146 * Includes ancestor txns' dirty pages not hidden by other txns'
1147 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1148 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1150 unsigned int mt_dirty_room;
1153 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1154 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1155 * raise this on a 64 bit machine.
1157 #define CURSOR_STACK 32
1161 /** Cursors are used for all DB operations.
1162 * A cursor holds a path of (page pointer, key index) from the DB
1163 * root to a position in the DB, plus other state. #MDB_DUPSORT
1164 * cursors include an xcursor to the current data item. Write txns
1165 * track their cursors and keep them up to date when data moves.
1166 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1167 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1170 /** Next cursor on this DB in this txn */
1171 MDB_cursor *mc_next;
1172 /** Backup of the original cursor if this cursor is a shadow */
1173 MDB_cursor *mc_backup;
1174 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1175 struct MDB_xcursor *mc_xcursor;
1176 /** The transaction that owns this cursor */
1178 /** The database handle this cursor operates on */
1180 /** The database record for this cursor */
1182 /** The database auxiliary record for this cursor */
1184 /** The @ref mt_dbflag for this database */
1185 unsigned char *mc_dbflag;
1186 unsigned short mc_snum; /**< number of pushed pages */
1187 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1188 /** @defgroup mdb_cursor Cursor Flags
1190 * Cursor state flags.
1193 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1194 #define C_EOF 0x02 /**< No more data */
1195 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1196 #define C_DEL 0x08 /**< last op was a cursor_del */
1197 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1199 unsigned int mc_flags; /**< @ref mdb_cursor */
1200 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1201 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1204 /** Context for sorted-dup records.
1205 * We could have gone to a fully recursive design, with arbitrarily
1206 * deep nesting of sub-databases. But for now we only handle these
1207 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1209 typedef struct MDB_xcursor {
1210 /** A sub-cursor for traversing the Dup DB */
1211 MDB_cursor mx_cursor;
1212 /** The database record for this Dup DB */
1214 /** The auxiliary DB record for this Dup DB */
1216 /** The @ref mt_dbflag for this Dup DB */
1217 unsigned char mx_dbflag;
1220 /** Check if there is an inited xcursor, so #XCURSOR_REFRESH() is proper */
1221 #define XCURSOR_INITED(mc) \
1222 ((mc)->mc_xcursor && ((mc)->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
1224 /** Update sub-page pointer, if any, in \b mc->mc_xcursor. Needed
1225 * when the node which contains the sub-page may have moved. Called
1226 * with \b mp = mc->mc_pg[mc->mc_top], \b ki = mc->mc_ki[mc->mc_top].
1228 #define XCURSOR_REFRESH(mc, mp, ki) do { \
1229 MDB_page *xr_pg = (mp); \
1230 MDB_node *xr_node = NODEPTR(xr_pg, ki); \
1231 if ((xr_node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) \
1232 (mc)->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(xr_node); \
1235 /** State of FreeDB old pages, stored in the MDB_env */
1236 typedef struct MDB_pgstate {
1237 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1238 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1241 /** The database environment. */
1243 HANDLE me_fd; /**< The main data file */
1244 HANDLE me_lfd; /**< The lock file */
1245 HANDLE me_mfd; /**< For writing and syncing the meta pages */
1246 /** Failed to update the meta page. Probably an I/O error. */
1247 #define MDB_FATAL_ERROR 0x80000000U
1248 /** Some fields are initialized. */
1249 #define MDB_ENV_ACTIVE 0x20000000U
1250 /** me_txkey is set */
1251 #define MDB_ENV_TXKEY 0x10000000U
1252 /** fdatasync is unreliable */
1253 #define MDB_FSYNCONLY 0x08000000U
1254 uint32_t me_flags; /**< @ref mdb_env */
1255 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1256 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1257 unsigned int me_maxreaders; /**< size of the reader table */
1258 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1259 volatile int me_close_readers;
1260 MDB_dbi me_numdbs; /**< number of DBs opened */
1261 MDB_dbi me_maxdbs; /**< size of the DB table */
1262 MDB_PID_T me_pid; /**< process ID of this env */
1263 char *me_path; /**< path to the DB files */
1264 char *me_map; /**< the memory map of the data file */
1265 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1266 MDB_meta *me_metas[NUM_METAS]; /**< pointers to the two meta pages */
1267 void *me_pbuf; /**< scratch area for DUPSORT put() */
1268 MDB_txn *me_txn; /**< current write transaction */
1269 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1270 size_t me_mapsize; /**< size of the data memory map */
1271 off_t me_size; /**< current file size */
1272 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1273 MDB_dbx *me_dbxs; /**< array of static DB info */
1274 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1275 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1276 pthread_key_t me_txkey; /**< thread-key for readers */
1277 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1278 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1279 # define me_pglast me_pgstate.mf_pglast
1280 # define me_pghead me_pgstate.mf_pghead
1281 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1282 /** IDL of pages that became unused in a write txn */
1283 MDB_IDL me_free_pgs;
1284 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1285 MDB_ID2L me_dirty_list;
1286 /** Max number of freelist items that can fit in a single overflow page */
1288 /** Max size of a node on a page */
1289 unsigned int me_nodemax;
1290 #if !(MDB_MAXKEYSIZE)
1291 unsigned int me_maxkey; /**< max size of a key */
1293 int me_live_reader; /**< have liveness lock in reader table */
1295 int me_pidquery; /**< Used in OpenProcess */
1297 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1298 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1299 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1301 mdb_mutex_t me_rmutex;
1302 mdb_mutex_t me_wmutex;
1304 void *me_userctx; /**< User-settable context */
1305 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1308 /** Nested transaction */
1309 typedef struct MDB_ntxn {
1310 MDB_txn mnt_txn; /**< the transaction */
1311 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1314 /** max number of pages to commit in one writev() call */
1315 #define MDB_COMMIT_PAGES 64
1316 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1317 #undef MDB_COMMIT_PAGES
1318 #define MDB_COMMIT_PAGES IOV_MAX
1321 /** max bytes to write in one call */
1322 #define MAX_WRITE (0x40000000U >> (sizeof(ssize_t) == 4))
1324 /** Check \b txn and \b dbi arguments to a function */
1325 #define TXN_DBI_EXIST(txn, dbi, validity) \
1326 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1328 /** Check for misused \b dbi handles */
1329 #define TXN_DBI_CHANGED(txn, dbi) \
1330 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1332 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1333 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1334 static int mdb_page_touch(MDB_cursor *mc);
1336 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
1337 "reset-tmp", "fail-begin", "fail-beginchild"}
1339 /* mdb_txn_end operation number, for logging */
1340 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
1341 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
1343 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */
1344 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */
1345 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */
1346 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */
1347 static void mdb_txn_end(MDB_txn *txn, unsigned mode);
1349 static int mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **mp, int *lvl);
1350 static int mdb_page_search_root(MDB_cursor *mc,
1351 MDB_val *key, int modify);
1352 #define MDB_PS_MODIFY 1
1353 #define MDB_PS_ROOTONLY 2
1354 #define MDB_PS_FIRST 4
1355 #define MDB_PS_LAST 8
1356 static int mdb_page_search(MDB_cursor *mc,
1357 MDB_val *key, int flags);
1358 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1360 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1361 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1362 pgno_t newpgno, unsigned int nflags);
1364 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1365 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1366 static int mdb_env_write_meta(MDB_txn *txn);
1367 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1368 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1370 static void mdb_env_close0(MDB_env *env, int excl);
1372 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1373 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1374 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1375 static void mdb_node_del(MDB_cursor *mc, int ksize);
1376 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1377 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft);
1378 static int mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data);
1379 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1380 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1382 static int mdb_rebalance(MDB_cursor *mc);
1383 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1385 static void mdb_cursor_pop(MDB_cursor *mc);
1386 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1388 static int mdb_cursor_del0(MDB_cursor *mc);
1389 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1390 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1391 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1392 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1393 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1395 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1396 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1398 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1399 static void mdb_xcursor_init0(MDB_cursor *mc);
1400 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1401 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force);
1403 static int mdb_drop0(MDB_cursor *mc, int subs);
1404 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1405 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1408 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1411 /** Compare two items pointing at size_t's of unknown alignment. */
1412 #ifdef MISALIGNED_OK
1413 # define mdb_cmp_clong mdb_cmp_long
1415 # define mdb_cmp_clong mdb_cmp_cint
1419 static SECURITY_DESCRIPTOR mdb_null_sd;
1420 static SECURITY_ATTRIBUTES mdb_all_sa;
1421 static int mdb_sec_inited;
1424 static int utf8_to_utf16(const char *src, struct MDB_name *dst, int xtra);
1427 /** Return the library version info. */
1429 mdb_version(int *major, int *minor, int *patch)
1431 if (major) *major = MDB_VERSION_MAJOR;
1432 if (minor) *minor = MDB_VERSION_MINOR;
1433 if (patch) *patch = MDB_VERSION_PATCH;
1434 return MDB_VERSION_STRING;
1437 /** Table of descriptions for LMDB @ref errors */
1438 static char *const mdb_errstr[] = {
1439 "MDB_KEYEXIST: Key/data pair already exists",
1440 "MDB_NOTFOUND: No matching key/data pair found",
1441 "MDB_PAGE_NOTFOUND: Requested page not found",
1442 "MDB_CORRUPTED: Located page was wrong type",
1443 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1444 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1445 "MDB_INVALID: File is not an LMDB file",
1446 "MDB_MAP_FULL: Environment mapsize limit reached",
1447 "MDB_DBS_FULL: Environment maxdbs limit reached",
1448 "MDB_READERS_FULL: Environment maxreaders limit reached",
1449 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1450 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1451 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1452 "MDB_PAGE_FULL: Internal error - page has no more space",
1453 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1454 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1455 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1456 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
1457 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1458 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1462 mdb_strerror(int err)
1465 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1466 * This works as long as no function between the call to mdb_strerror
1467 * and the actual use of the message uses more than 4K of stack.
1469 #define MSGSIZE 1024
1470 #define PADSIZE 4096
1471 char buf[MSGSIZE+PADSIZE], *ptr = buf;
1475 return ("Successful return: 0");
1477 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1478 i = err - MDB_KEYEXIST;
1479 return mdb_errstr[i];
1483 /* These are the C-runtime error codes we use. The comment indicates
1484 * their numeric value, and the Win32 error they would correspond to
1485 * if the error actually came from a Win32 API. A major mess, we should
1486 * have used LMDB-specific error codes for everything.
1489 case ENOENT: /* 2, FILE_NOT_FOUND */
1490 case EIO: /* 5, ACCESS_DENIED */
1491 case ENOMEM: /* 12, INVALID_ACCESS */
1492 case EACCES: /* 13, INVALID_DATA */
1493 case EBUSY: /* 16, CURRENT_DIRECTORY */
1494 case EINVAL: /* 22, BAD_COMMAND */
1495 case ENOSPC: /* 28, OUT_OF_PAPER */
1496 return strerror(err);
1501 FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM |
1502 FORMAT_MESSAGE_IGNORE_INSERTS,
1503 NULL, err, 0, ptr, MSGSIZE, (va_list *)buf+MSGSIZE);
1506 return strerror(err);
1510 /** assert(3) variant in cursor context */
1511 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1512 /** assert(3) variant in transaction context */
1513 #define mdb_tassert(txn, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1514 /** assert(3) variant in environment context */
1515 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1518 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1519 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1522 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1523 const char *func, const char *file, int line)
1526 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1527 file, line, expr_txt, func);
1528 if (env->me_assert_func)
1529 env->me_assert_func(env, buf);
1530 fprintf(stderr, "%s\n", buf);
1534 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1538 /** Return the page number of \b mp which may be sub-page, for debug output */
1540 mdb_dbg_pgno(MDB_page *mp)
1543 COPY_PGNO(ret, mp->mp_pgno);
1547 /** Display a key in hexadecimal and return the address of the result.
1548 * @param[in] key the key to display
1549 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1550 * @return The key in hexadecimal form.
1553 mdb_dkey(MDB_val *key, char *buf)
1556 unsigned char *c = key->mv_data;
1562 if (key->mv_size > DKBUF_MAXKEYSIZE)
1563 return "MDB_MAXKEYSIZE";
1564 /* may want to make this a dynamic check: if the key is mostly
1565 * printable characters, print it as-is instead of converting to hex.
1569 for (i=0; i<key->mv_size; i++)
1570 ptr += sprintf(ptr, "%02x", *c++);
1572 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1578 mdb_leafnode_type(MDB_node *n)
1580 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1581 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1582 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1585 /** Display all the keys in the page. */
1587 mdb_page_list(MDB_page *mp)
1589 pgno_t pgno = mdb_dbg_pgno(mp);
1590 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1592 unsigned int i, nkeys, nsize, total = 0;
1596 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1597 case P_BRANCH: type = "Branch page"; break;
1598 case P_LEAF: type = "Leaf page"; break;
1599 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1600 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1601 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1603 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1604 pgno, mp->mp_pages, state);
1607 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1608 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1611 fprintf(stderr, "Bad page %"Z"u flags 0x%X\n", pgno, mp->mp_flags);
1615 nkeys = NUMKEYS(mp);
1616 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1618 for (i=0; i<nkeys; i++) {
1619 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1620 key.mv_size = nsize = mp->mp_pad;
1621 key.mv_data = LEAF2KEY(mp, i, nsize);
1623 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1626 node = NODEPTR(mp, i);
1627 key.mv_size = node->mn_ksize;
1628 key.mv_data = node->mn_data;
1629 nsize = NODESIZE + key.mv_size;
1630 if (IS_BRANCH(mp)) {
1631 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1635 if (F_ISSET(node->mn_flags, F_BIGDATA))
1636 nsize += sizeof(pgno_t);
1638 nsize += NODEDSZ(node);
1640 nsize += sizeof(indx_t);
1641 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1642 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1644 total = EVEN(total);
1646 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1647 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1651 mdb_cursor_chk(MDB_cursor *mc)
1657 if (!mc->mc_snum || !(mc->mc_flags & C_INITIALIZED)) return;
1658 for (i=0; i<mc->mc_top; i++) {
1660 node = NODEPTR(mp, mc->mc_ki[i]);
1661 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1664 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1666 if (XCURSOR_INITED(mc)) {
1667 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
1668 if (((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) &&
1669 mc->mc_xcursor->mx_cursor.mc_pg[0] != NODEDATA(node)) {
1677 /** Count all the pages in each DB and in the freelist
1678 * and make sure it matches the actual number of pages
1680 * All named DBs must be open for a correct count.
1682 static void mdb_audit(MDB_txn *txn)
1686 MDB_ID freecount, count;
1691 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1692 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1693 freecount += *(MDB_ID *)data.mv_data;
1694 mdb_tassert(txn, rc == MDB_NOTFOUND);
1697 for (i = 0; i<txn->mt_numdbs; i++) {
1699 if (!(txn->mt_dbflags[i] & DB_VALID))
1701 mdb_cursor_init(&mc, txn, i, &mx);
1702 if (txn->mt_dbs[i].md_root == P_INVALID)
1704 count += txn->mt_dbs[i].md_branch_pages +
1705 txn->mt_dbs[i].md_leaf_pages +
1706 txn->mt_dbs[i].md_overflow_pages;
1707 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1708 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1709 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1712 mp = mc.mc_pg[mc.mc_top];
1713 for (j=0; j<NUMKEYS(mp); j++) {
1714 MDB_node *leaf = NODEPTR(mp, j);
1715 if (leaf->mn_flags & F_SUBDATA) {
1717 memcpy(&db, NODEDATA(leaf), sizeof(db));
1718 count += db.md_branch_pages + db.md_leaf_pages +
1719 db.md_overflow_pages;
1723 mdb_tassert(txn, rc == MDB_NOTFOUND);
1726 if (freecount + count + NUM_METAS != txn->mt_next_pgno) {
1727 fprintf(stderr, "audit: %"Z"u freecount: %"Z"u count: %"Z"u total: %"Z"u next_pgno: %"Z"u\n",
1728 txn->mt_txnid, freecount, count+NUM_METAS,
1729 freecount+count+NUM_METAS, txn->mt_next_pgno);
1735 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1737 return txn->mt_dbxs[dbi].md_cmp(a, b);
1741 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1743 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1744 #if UINT_MAX < SIZE_MAX
1745 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1746 dcmp = mdb_cmp_clong;
1751 /** Allocate memory for a page.
1752 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1753 * Set #MDB_TXN_ERROR on failure.
1756 mdb_page_malloc(MDB_txn *txn, unsigned num)
1758 MDB_env *env = txn->mt_env;
1759 MDB_page *ret = env->me_dpages;
1760 size_t psize = env->me_psize, sz = psize, off;
1761 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1762 * For a single page alloc, we init everything after the page header.
1763 * For multi-page, we init the final page; if the caller needed that
1764 * many pages they will be filling in at least up to the last page.
1768 VGMEMP_ALLOC(env, ret, sz);
1769 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1770 env->me_dpages = ret->mp_next;
1773 psize -= off = PAGEHDRSZ;
1778 if ((ret = malloc(sz)) != NULL) {
1779 VGMEMP_ALLOC(env, ret, sz);
1780 if (!(env->me_flags & MDB_NOMEMINIT)) {
1781 memset((char *)ret + off, 0, psize);
1785 txn->mt_flags |= MDB_TXN_ERROR;
1789 /** Free a single page.
1790 * Saves single pages to a list, for future reuse.
1791 * (This is not used for multi-page overflow pages.)
1794 mdb_page_free(MDB_env *env, MDB_page *mp)
1796 mp->mp_next = env->me_dpages;
1797 VGMEMP_FREE(env, mp);
1798 env->me_dpages = mp;
1801 /** Free a dirty page */
1803 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1805 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1806 mdb_page_free(env, dp);
1808 /* large pages just get freed directly */
1809 VGMEMP_FREE(env, dp);
1814 /** Return all dirty pages to dpage list */
1816 mdb_dlist_free(MDB_txn *txn)
1818 MDB_env *env = txn->mt_env;
1819 MDB_ID2L dl = txn->mt_u.dirty_list;
1820 unsigned i, n = dl[0].mid;
1822 for (i = 1; i <= n; i++) {
1823 mdb_dpage_free(env, dl[i].mptr);
1828 /** Loosen or free a single page.
1829 * Saves single pages to a list for future reuse
1830 * in this same txn. It has been pulled from the freeDB
1831 * and already resides on the dirty list, but has been
1832 * deleted. Use these pages first before pulling again
1835 * If the page wasn't dirtied in this txn, just add it
1836 * to this txn's free list.
1839 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1842 pgno_t pgno = mp->mp_pgno;
1843 MDB_txn *txn = mc->mc_txn;
1845 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1846 if (txn->mt_parent) {
1847 MDB_ID2 *dl = txn->mt_u.dirty_list;
1848 /* If txn has a parent, make sure the page is in our
1852 unsigned x = mdb_mid2l_search(dl, pgno);
1853 if (x <= dl[0].mid && dl[x].mid == pgno) {
1854 if (mp != dl[x].mptr) { /* bad cursor? */
1855 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1856 txn->mt_flags |= MDB_TXN_ERROR;
1857 return MDB_CORRUPTED;
1864 /* no parent txn, so it's just ours */
1869 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1871 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1872 txn->mt_loose_pgs = mp;
1873 txn->mt_loose_count++;
1874 mp->mp_flags |= P_LOOSE;
1876 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1884 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1885 * @param[in] mc A cursor handle for the current operation.
1886 * @param[in] pflags Flags of the pages to update:
1887 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1888 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1889 * @return 0 on success, non-zero on failure.
1892 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1894 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1895 MDB_txn *txn = mc->mc_txn;
1896 MDB_cursor *m3, *m0 = mc;
1901 int rc = MDB_SUCCESS, level;
1903 /* Mark pages seen by cursors */
1904 if (mc->mc_flags & C_UNTRACK)
1905 mc = NULL; /* will find mc in mt_cursors */
1906 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1907 for (; mc; mc=mc->mc_next) {
1908 if (!(mc->mc_flags & C_INITIALIZED))
1910 for (m3 = mc;; m3 = &mx->mx_cursor) {
1912 for (j=0; j<m3->mc_snum; j++) {
1914 if ((mp->mp_flags & Mask) == pflags)
1915 mp->mp_flags ^= P_KEEP;
1917 mx = m3->mc_xcursor;
1918 /* Proceed to mx if it is at a sub-database */
1919 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1921 if (! (mp && (mp->mp_flags & P_LEAF)))
1923 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1924 if (!(leaf->mn_flags & F_SUBDATA))
1933 /* Mark dirty root pages */
1934 for (i=0; i<txn->mt_numdbs; i++) {
1935 if (txn->mt_dbflags[i] & DB_DIRTY) {
1936 pgno_t pgno = txn->mt_dbs[i].md_root;
1937 if (pgno == P_INVALID)
1939 if ((rc = mdb_page_get(m0, pgno, &dp, &level)) != MDB_SUCCESS)
1941 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1942 dp->mp_flags ^= P_KEEP;
1950 static int mdb_page_flush(MDB_txn *txn, int keep);
1952 /** Spill pages from the dirty list back to disk.
1953 * This is intended to prevent running into #MDB_TXN_FULL situations,
1954 * but note that they may still occur in a few cases:
1955 * 1) our estimate of the txn size could be too small. Currently this
1956 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1957 * 2) child txns may run out of space if their parents dirtied a
1958 * lot of pages and never spilled them. TODO: we probably should do
1959 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1960 * the parent's dirty_room is below a given threshold.
1962 * Otherwise, if not using nested txns, it is expected that apps will
1963 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1964 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1965 * If the txn never references them again, they can be left alone.
1966 * If the txn only reads them, they can be used without any fuss.
1967 * If the txn writes them again, they can be dirtied immediately without
1968 * going thru all of the work of #mdb_page_touch(). Such references are
1969 * handled by #mdb_page_unspill().
1971 * Also note, we never spill DB root pages, nor pages of active cursors,
1972 * because we'll need these back again soon anyway. And in nested txns,
1973 * we can't spill a page in a child txn if it was already spilled in a
1974 * parent txn. That would alter the parent txns' data even though
1975 * the child hasn't committed yet, and we'd have no way to undo it if
1976 * the child aborted.
1978 * @param[in] m0 cursor A cursor handle identifying the transaction and
1979 * database for which we are checking space.
1980 * @param[in] key For a put operation, the key being stored.
1981 * @param[in] data For a put operation, the data being stored.
1982 * @return 0 on success, non-zero on failure.
1985 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1987 MDB_txn *txn = m0->mc_txn;
1989 MDB_ID2L dl = txn->mt_u.dirty_list;
1990 unsigned int i, j, need;
1993 if (m0->mc_flags & C_SUB)
1996 /* Estimate how much space this op will take */
1997 i = m0->mc_db->md_depth;
1998 /* Named DBs also dirty the main DB */
1999 if (m0->mc_dbi >= CORE_DBS)
2000 i += txn->mt_dbs[MAIN_DBI].md_depth;
2001 /* For puts, roughly factor in the key+data size */
2003 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
2004 i += i; /* double it for good measure */
2007 if (txn->mt_dirty_room > i)
2010 if (!txn->mt_spill_pgs) {
2011 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
2012 if (!txn->mt_spill_pgs)
2015 /* purge deleted slots */
2016 MDB_IDL sl = txn->mt_spill_pgs;
2017 unsigned int num = sl[0];
2019 for (i=1; i<=num; i++) {
2026 /* Preserve pages which may soon be dirtied again */
2027 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
2030 /* Less aggressive spill - we originally spilled the entire dirty list,
2031 * with a few exceptions for cursor pages and DB root pages. But this
2032 * turns out to be a lot of wasted effort because in a large txn many
2033 * of those pages will need to be used again. So now we spill only 1/8th
2034 * of the dirty pages. Testing revealed this to be a good tradeoff,
2035 * better than 1/2, 1/4, or 1/10.
2037 if (need < MDB_IDL_UM_MAX / 8)
2038 need = MDB_IDL_UM_MAX / 8;
2040 /* Save the page IDs of all the pages we're flushing */
2041 /* flush from the tail forward, this saves a lot of shifting later on. */
2042 for (i=dl[0].mid; i && need; i--) {
2043 MDB_ID pn = dl[i].mid << 1;
2045 if (dp->mp_flags & (P_LOOSE|P_KEEP))
2047 /* Can't spill twice, make sure it's not already in a parent's
2050 if (txn->mt_parent) {
2052 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
2053 if (tx2->mt_spill_pgs) {
2054 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
2055 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
2056 dp->mp_flags |= P_KEEP;
2064 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2068 mdb_midl_sort(txn->mt_spill_pgs);
2070 /* Flush the spilled part of dirty list */
2071 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2074 /* Reset any dirty pages we kept that page_flush didn't see */
2075 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2078 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2082 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2084 mdb_find_oldest(MDB_txn *txn)
2087 txnid_t mr, oldest = txn->mt_txnid - 1;
2088 if (txn->mt_env->me_txns) {
2089 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2090 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2101 /** Add a page to the txn's dirty list */
2103 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2106 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2108 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2109 insert = mdb_mid2l_append;
2111 insert = mdb_mid2l_insert;
2113 mid.mid = mp->mp_pgno;
2115 rc = insert(txn->mt_u.dirty_list, &mid);
2116 mdb_tassert(txn, rc == 0);
2117 txn->mt_dirty_room--;
2120 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2121 * me_pghead and mt_next_pgno. Set #MDB_TXN_ERROR on failure.
2123 * If there are free pages available from older transactions, they
2124 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2125 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2126 * and move me_pglast to say which records were consumed. Only this
2127 * function can create me_pghead and move me_pglast/mt_next_pgno.
2128 * @param[in] mc cursor A cursor handle identifying the transaction and
2129 * database for which we are allocating.
2130 * @param[in] num the number of pages to allocate.
2131 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2132 * will always be satisfied by a single contiguous chunk of memory.
2133 * @return 0 on success, non-zero on failure.
2136 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2138 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2139 /* Get at most <Max_retries> more freeDB records once me_pghead
2140 * has enough pages. If not enough, use new pages from the map.
2141 * If <Paranoid> and mc is updating the freeDB, only get new
2142 * records if me_pghead is empty. Then the freelist cannot play
2143 * catch-up with itself by growing while trying to save it.
2145 enum { Paranoid = 1, Max_retries = 500 };
2147 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2149 int rc, retry = num * 60;
2150 MDB_txn *txn = mc->mc_txn;
2151 MDB_env *env = txn->mt_env;
2152 pgno_t pgno, *mop = env->me_pghead;
2153 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2155 txnid_t oldest = 0, last;
2160 /* If there are any loose pages, just use them */
2161 if (num == 1 && txn->mt_loose_pgs) {
2162 np = txn->mt_loose_pgs;
2163 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2164 txn->mt_loose_count--;
2165 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2173 /* If our dirty list is already full, we can't do anything */
2174 if (txn->mt_dirty_room == 0) {
2179 for (op = MDB_FIRST;; op = MDB_NEXT) {
2184 /* Seek a big enough contiguous page range. Prefer
2185 * pages at the tail, just truncating the list.
2191 if (mop[i-n2] == pgno+n2)
2198 if (op == MDB_FIRST) { /* 1st iteration */
2199 /* Prepare to fetch more and coalesce */
2200 last = env->me_pglast;
2201 oldest = env->me_pgoldest;
2202 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2205 key.mv_data = &last; /* will look up last+1 */
2206 key.mv_size = sizeof(last);
2208 if (Paranoid && mc->mc_dbi == FREE_DBI)
2211 if (Paranoid && retry < 0 && mop_len)
2215 /* Do not fetch more if the record will be too recent */
2216 if (oldest <= last) {
2218 oldest = mdb_find_oldest(txn);
2219 env->me_pgoldest = oldest;
2225 rc = mdb_cursor_get(&m2, &key, NULL, op);
2227 if (rc == MDB_NOTFOUND)
2231 last = *(txnid_t*)key.mv_data;
2232 if (oldest <= last) {
2234 oldest = mdb_find_oldest(txn);
2235 env->me_pgoldest = oldest;
2241 np = m2.mc_pg[m2.mc_top];
2242 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2243 if ((rc = mdb_node_read(&m2, leaf, &data)) != MDB_SUCCESS)
2246 idl = (MDB_ID *) data.mv_data;
2249 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2254 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2256 mop = env->me_pghead;
2258 env->me_pglast = last;
2260 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2261 last, txn->mt_dbs[FREE_DBI].md_root, i));
2263 DPRINTF(("IDL %"Z"u", idl[j]));
2265 /* Merge in descending sorted order */
2266 mdb_midl_xmerge(mop, idl);
2270 /* Use new pages from the map when nothing suitable in the freeDB */
2272 pgno = txn->mt_next_pgno;
2273 if (pgno + num >= env->me_maxpg) {
2274 DPUTS("DB size maxed out");
2280 if (env->me_flags & MDB_WRITEMAP) {
2281 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2283 if (!(np = mdb_page_malloc(txn, num))) {
2289 mop[0] = mop_len -= num;
2290 /* Move any stragglers down */
2291 for (j = i-num; j < mop_len; )
2292 mop[++j] = mop[++i];
2294 txn->mt_next_pgno = pgno + num;
2297 mdb_page_dirty(txn, np);
2303 txn->mt_flags |= MDB_TXN_ERROR;
2307 /** Copy the used portions of a non-overflow page.
2308 * @param[in] dst page to copy into
2309 * @param[in] src page to copy from
2310 * @param[in] psize size of a page
2313 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2315 enum { Align = sizeof(pgno_t) };
2316 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2318 /* If page isn't full, just copy the used portion. Adjust
2319 * alignment so memcpy may copy words instead of bytes.
2321 if ((unused &= -Align) && !IS_LEAF2(src)) {
2322 upper = (upper + PAGEBASE) & -Align;
2323 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2324 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2327 memcpy(dst, src, psize - unused);
2331 /** Pull a page off the txn's spill list, if present.
2332 * If a page being referenced was spilled to disk in this txn, bring
2333 * it back and make it dirty/writable again.
2334 * @param[in] txn the transaction handle.
2335 * @param[in] mp the page being referenced. It must not be dirty.
2336 * @param[out] ret the writable page, if any. ret is unchanged if
2337 * mp wasn't spilled.
2340 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2342 MDB_env *env = txn->mt_env;
2345 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2347 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2348 if (!tx2->mt_spill_pgs)
2350 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2351 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2354 if (txn->mt_dirty_room == 0)
2355 return MDB_TXN_FULL;
2356 if (IS_OVERFLOW(mp))
2360 if (env->me_flags & MDB_WRITEMAP) {
2363 np = mdb_page_malloc(txn, num);
2367 memcpy(np, mp, num * env->me_psize);
2369 mdb_page_copy(np, mp, env->me_psize);
2372 /* If in current txn, this page is no longer spilled.
2373 * If it happens to be the last page, truncate the spill list.
2374 * Otherwise mark it as deleted by setting the LSB.
2376 if (x == txn->mt_spill_pgs[0])
2377 txn->mt_spill_pgs[0]--;
2379 txn->mt_spill_pgs[x] |= 1;
2380 } /* otherwise, if belonging to a parent txn, the
2381 * page remains spilled until child commits
2384 mdb_page_dirty(txn, np);
2385 np->mp_flags |= P_DIRTY;
2393 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2394 * Set #MDB_TXN_ERROR on failure.
2395 * @param[in] mc cursor pointing to the page to be touched
2396 * @return 0 on success, non-zero on failure.
2399 mdb_page_touch(MDB_cursor *mc)
2401 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2402 MDB_txn *txn = mc->mc_txn;
2403 MDB_cursor *m2, *m3;
2407 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2408 if (txn->mt_flags & MDB_TXN_SPILLS) {
2410 rc = mdb_page_unspill(txn, mp, &np);
2416 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2417 (rc = mdb_page_alloc(mc, 1, &np)))
2420 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2421 mp->mp_pgno, pgno));
2422 mdb_cassert(mc, mp->mp_pgno != pgno);
2423 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2424 /* Update the parent page, if any, to point to the new page */
2426 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2427 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2428 SETPGNO(node, pgno);
2430 mc->mc_db->md_root = pgno;
2432 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2433 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2435 /* If txn has a parent, make sure the page is in our
2439 unsigned x = mdb_mid2l_search(dl, pgno);
2440 if (x <= dl[0].mid && dl[x].mid == pgno) {
2441 if (mp != dl[x].mptr) { /* bad cursor? */
2442 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2443 txn->mt_flags |= MDB_TXN_ERROR;
2444 return MDB_CORRUPTED;
2449 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2451 np = mdb_page_malloc(txn, 1);
2456 rc = mdb_mid2l_insert(dl, &mid);
2457 mdb_cassert(mc, rc == 0);
2462 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2464 np->mp_flags |= P_DIRTY;
2467 /* Adjust cursors pointing to mp */
2468 mc->mc_pg[mc->mc_top] = np;
2469 m2 = txn->mt_cursors[mc->mc_dbi];
2470 if (mc->mc_flags & C_SUB) {
2471 for (; m2; m2=m2->mc_next) {
2472 m3 = &m2->mc_xcursor->mx_cursor;
2473 if (m3->mc_snum < mc->mc_snum) continue;
2474 if (m3->mc_pg[mc->mc_top] == mp)
2475 m3->mc_pg[mc->mc_top] = np;
2478 for (; m2; m2=m2->mc_next) {
2479 if (m2->mc_snum < mc->mc_snum) continue;
2480 if (m2 == mc) continue;
2481 if (m2->mc_pg[mc->mc_top] == mp) {
2482 m2->mc_pg[mc->mc_top] = np;
2483 if (XCURSOR_INITED(m2) && IS_LEAF(np))
2484 XCURSOR_REFRESH(m2, np, m2->mc_ki[mc->mc_top]);
2491 txn->mt_flags |= MDB_TXN_ERROR;
2496 mdb_env_sync(MDB_env *env, int force)
2499 if (env->me_flags & MDB_RDONLY)
2501 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2502 if (env->me_flags & MDB_WRITEMAP) {
2503 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2504 ? MS_ASYNC : MS_SYNC;
2505 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2508 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2512 #ifdef BROKEN_FDATASYNC
2513 if (env->me_flags & MDB_FSYNCONLY) {
2514 if (fsync(env->me_fd))
2518 if (MDB_FDATASYNC(env->me_fd))
2525 /** Back up parent txn's cursors, then grab the originals for tracking */
2527 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2529 MDB_cursor *mc, *bk;
2534 for (i = src->mt_numdbs; --i >= 0; ) {
2535 if ((mc = src->mt_cursors[i]) != NULL) {
2536 size = sizeof(MDB_cursor);
2538 size += sizeof(MDB_xcursor);
2539 for (; mc; mc = bk->mc_next) {
2545 mc->mc_db = &dst->mt_dbs[i];
2546 /* Kill pointers into src to reduce abuse: The
2547 * user may not use mc until dst ends. But we need a valid
2548 * txn pointer here for cursor fixups to keep working.
2551 mc->mc_dbflag = &dst->mt_dbflags[i];
2552 if ((mx = mc->mc_xcursor) != NULL) {
2553 *(MDB_xcursor *)(bk+1) = *mx;
2554 mx->mx_cursor.mc_txn = dst;
2556 mc->mc_next = dst->mt_cursors[i];
2557 dst->mt_cursors[i] = mc;
2564 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2565 * @param[in] txn the transaction handle.
2566 * @param[in] merge true to keep changes to parent cursors, false to revert.
2567 * @return 0 on success, non-zero on failure.
2570 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2572 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2576 for (i = txn->mt_numdbs; --i >= 0; ) {
2577 for (mc = cursors[i]; mc; mc = next) {
2579 if ((bk = mc->mc_backup) != NULL) {
2581 /* Commit changes to parent txn */
2582 mc->mc_next = bk->mc_next;
2583 mc->mc_backup = bk->mc_backup;
2584 mc->mc_txn = bk->mc_txn;
2585 mc->mc_db = bk->mc_db;
2586 mc->mc_dbflag = bk->mc_dbflag;
2587 if ((mx = mc->mc_xcursor) != NULL)
2588 mx->mx_cursor.mc_txn = bk->mc_txn;
2590 /* Abort nested txn */
2592 if ((mx = mc->mc_xcursor) != NULL)
2593 *mx = *(MDB_xcursor *)(bk+1);
2597 /* Only malloced cursors are permanently tracked. */
2604 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2610 Pidset = F_SETLK, Pidcheck = F_GETLK
2614 /** Set or check a pid lock. Set returns 0 on success.
2615 * Check returns 0 if the process is certainly dead, nonzero if it may
2616 * be alive (the lock exists or an error happened so we do not know).
2618 * On Windows Pidset is a no-op, we merely check for the existence
2619 * of the process with the given pid. On POSIX we use a single byte
2620 * lock on the lockfile, set at an offset equal to the pid.
2623 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2625 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2628 if (op == Pidcheck) {
2629 h = OpenProcess(env->me_pidquery, FALSE, pid);
2630 /* No documented "no such process" code, but other program use this: */
2632 return ErrCode() != ERROR_INVALID_PARAMETER;
2633 /* A process exists until all handles to it close. Has it exited? */
2634 ret = WaitForSingleObject(h, 0) != 0;
2641 struct flock lock_info;
2642 memset(&lock_info, 0, sizeof(lock_info));
2643 lock_info.l_type = F_WRLCK;
2644 lock_info.l_whence = SEEK_SET;
2645 lock_info.l_start = pid;
2646 lock_info.l_len = 1;
2647 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2648 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2650 } else if ((rc = ErrCode()) == EINTR) {
2658 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2659 * @param[in] txn the transaction handle to initialize
2660 * @return 0 on success, non-zero on failure.
2663 mdb_txn_renew0(MDB_txn *txn)
2665 MDB_env *env = txn->mt_env;
2666 MDB_txninfo *ti = env->me_txns;
2668 unsigned int i, nr, flags = txn->mt_flags;
2670 int rc, new_notls = 0;
2672 if ((flags &= MDB_TXN_RDONLY) != 0) {
2674 meta = mdb_env_pick_meta(env);
2675 txn->mt_txnid = meta->mm_txnid;
2676 txn->mt_u.reader = NULL;
2678 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2679 pthread_getspecific(env->me_txkey);
2681 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2682 return MDB_BAD_RSLOT;
2684 MDB_PID_T pid = env->me_pid;
2685 MDB_THR_T tid = pthread_self();
2686 mdb_mutexref_t rmutex = env->me_rmutex;
2688 if (!env->me_live_reader) {
2689 rc = mdb_reader_pid(env, Pidset, pid);
2692 env->me_live_reader = 1;
2695 if (LOCK_MUTEX(rc, env, rmutex))
2697 nr = ti->mti_numreaders;
2698 for (i=0; i<nr; i++)
2699 if (ti->mti_readers[i].mr_pid == 0)
2701 if (i == env->me_maxreaders) {
2702 UNLOCK_MUTEX(rmutex);
2703 return MDB_READERS_FULL;
2705 r = &ti->mti_readers[i];
2706 /* Claim the reader slot, carefully since other code
2707 * uses the reader table un-mutexed: First reset the
2708 * slot, next publish it in mti_numreaders. After
2709 * that, it is safe for mdb_env_close() to touch it.
2710 * When it will be closed, we can finally claim it.
2713 r->mr_txnid = (txnid_t)-1;
2716 ti->mti_numreaders = ++nr;
2717 env->me_close_readers = nr;
2719 UNLOCK_MUTEX(rmutex);
2721 new_notls = (env->me_flags & MDB_NOTLS);
2722 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2727 do /* LY: Retry on a race, ITS#7970. */
2728 r->mr_txnid = ti->mti_txnid;
2729 while(r->mr_txnid != ti->mti_txnid);
2730 txn->mt_txnid = r->mr_txnid;
2731 txn->mt_u.reader = r;
2732 meta = env->me_metas[txn->mt_txnid & 1];
2736 /* Not yet touching txn == env->me_txn0, it may be active */
2738 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2740 txn->mt_txnid = ti->mti_txnid;
2741 meta = env->me_metas[txn->mt_txnid & 1];
2743 meta = mdb_env_pick_meta(env);
2744 txn->mt_txnid = meta->mm_txnid;
2748 if (txn->mt_txnid == mdb_debug_start)
2751 txn->mt_child = NULL;
2752 txn->mt_loose_pgs = NULL;
2753 txn->mt_loose_count = 0;
2754 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2755 txn->mt_u.dirty_list = env->me_dirty_list;
2756 txn->mt_u.dirty_list[0].mid = 0;
2757 txn->mt_free_pgs = env->me_free_pgs;
2758 txn->mt_free_pgs[0] = 0;
2759 txn->mt_spill_pgs = NULL;
2761 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2764 /* Copy the DB info and flags */
2765 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
2767 /* Moved to here to avoid a data race in read TXNs */
2768 txn->mt_next_pgno = meta->mm_last_pg+1;
2770 txn->mt_flags = flags;
2773 txn->mt_numdbs = env->me_numdbs;
2774 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
2775 x = env->me_dbflags[i];
2776 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2777 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2779 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2780 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2782 if (env->me_flags & MDB_FATAL_ERROR) {
2783 DPUTS("environment had fatal error, must shutdown!");
2785 } else if (env->me_maxpg < txn->mt_next_pgno) {
2786 rc = MDB_MAP_RESIZED;
2790 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
2795 mdb_txn_renew(MDB_txn *txn)
2799 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED))
2802 rc = mdb_txn_renew0(txn);
2803 if (rc == MDB_SUCCESS) {
2804 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2805 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2806 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2812 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2816 int rc, size, tsize;
2818 flags &= MDB_TXN_BEGIN_FLAGS;
2819 flags |= env->me_flags & MDB_WRITEMAP;
2821 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2825 /* Nested transactions: Max 1 child, write txns only, no writemap */
2826 flags |= parent->mt_flags;
2827 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
2828 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2830 /* Child txns save MDB_pgstate and use own copy of cursors */
2831 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
2832 size += tsize = sizeof(MDB_ntxn);
2833 } else if (flags & MDB_RDONLY) {
2834 size = env->me_maxdbs * (sizeof(MDB_db)+1);
2835 size += tsize = sizeof(MDB_txn);
2837 /* Reuse preallocated write txn. However, do not touch it until
2838 * mdb_txn_renew0() succeeds, since it currently may be active.
2843 if ((txn = calloc(1, size)) == NULL) {
2844 DPRINTF(("calloc: %s", strerror(errno)));
2847 txn->mt_dbxs = env->me_dbxs; /* static */
2848 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2849 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
2850 txn->mt_flags = flags;
2855 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2856 txn->mt_dbiseqs = parent->mt_dbiseqs;
2857 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2858 if (!txn->mt_u.dirty_list ||
2859 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2861 free(txn->mt_u.dirty_list);
2865 txn->mt_txnid = parent->mt_txnid;
2866 txn->mt_dirty_room = parent->mt_dirty_room;
2867 txn->mt_u.dirty_list[0].mid = 0;
2868 txn->mt_spill_pgs = NULL;
2869 txn->mt_next_pgno = parent->mt_next_pgno;
2870 parent->mt_flags |= MDB_TXN_HAS_CHILD;
2871 parent->mt_child = txn;
2872 txn->mt_parent = parent;
2873 txn->mt_numdbs = parent->mt_numdbs;
2874 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2875 /* Copy parent's mt_dbflags, but clear DB_NEW */
2876 for (i=0; i<txn->mt_numdbs; i++)
2877 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2879 ntxn = (MDB_ntxn *)txn;
2880 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2881 if (env->me_pghead) {
2882 size = MDB_IDL_SIZEOF(env->me_pghead);
2883 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2885 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2890 rc = mdb_cursor_shadow(parent, txn);
2892 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
2893 } else { /* MDB_RDONLY */
2894 txn->mt_dbiseqs = env->me_dbiseqs;
2896 rc = mdb_txn_renew0(txn);
2899 if (txn != env->me_txn0)
2902 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
2904 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2905 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2906 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2913 mdb_txn_env(MDB_txn *txn)
2915 if(!txn) return NULL;
2920 mdb_txn_id(MDB_txn *txn)
2923 return txn->mt_txnid;
2926 /** Export or close DBI handles opened in this txn. */
2928 mdb_dbis_update(MDB_txn *txn, int keep)
2931 MDB_dbi n = txn->mt_numdbs;
2932 MDB_env *env = txn->mt_env;
2933 unsigned char *tdbflags = txn->mt_dbflags;
2935 for (i = n; --i >= CORE_DBS;) {
2936 if (tdbflags[i] & DB_NEW) {
2938 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2940 char *ptr = env->me_dbxs[i].md_name.mv_data;
2942 env->me_dbxs[i].md_name.mv_data = NULL;
2943 env->me_dbxs[i].md_name.mv_size = 0;
2944 env->me_dbflags[i] = 0;
2945 env->me_dbiseqs[i]++;
2951 if (keep && env->me_numdbs < n)
2955 /** End a transaction, except successful commit of a nested transaction.
2956 * May be called twice for readonly txns: First reset it, then abort.
2957 * @param[in] txn the transaction handle to end
2958 * @param[in] mode why and how to end the transaction
2961 mdb_txn_end(MDB_txn *txn, unsigned mode)
2963 MDB_env *env = txn->mt_env;
2965 static const char *const names[] = MDB_END_NAMES;
2968 /* Export or close DBI handles opened in this txn */
2969 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
2971 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2972 names[mode & MDB_END_OPMASK],
2973 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2974 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2976 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2977 if (txn->mt_u.reader) {
2978 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2979 if (!(env->me_flags & MDB_NOTLS)) {
2980 txn->mt_u.reader = NULL; /* txn does not own reader */
2981 } else if (mode & MDB_END_SLOT) {
2982 txn->mt_u.reader->mr_pid = 0;
2983 txn->mt_u.reader = NULL;
2984 } /* else txn owns the slot until it does MDB_END_SLOT */
2986 txn->mt_numdbs = 0; /* prevent further DBI activity */
2987 txn->mt_flags |= MDB_TXN_FINISHED;
2989 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
2990 pgno_t *pghead = env->me_pghead;
2992 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
2993 mdb_cursors_close(txn, 0);
2994 if (!(env->me_flags & MDB_WRITEMAP)) {
2995 mdb_dlist_free(txn);
2999 txn->mt_flags = MDB_TXN_FINISHED;
3001 if (!txn->mt_parent) {
3002 mdb_midl_shrink(&txn->mt_free_pgs);
3003 env->me_free_pgs = txn->mt_free_pgs;
3005 env->me_pghead = NULL;
3009 mode = 0; /* txn == env->me_txn0, do not free() it */
3011 /* The writer mutex was locked in mdb_txn_begin. */
3013 UNLOCK_MUTEX(env->me_wmutex);
3015 txn->mt_parent->mt_child = NULL;
3016 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
3017 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
3018 mdb_midl_free(txn->mt_free_pgs);
3019 mdb_midl_free(txn->mt_spill_pgs);
3020 free(txn->mt_u.dirty_list);
3023 mdb_midl_free(pghead);
3026 if (mode & MDB_END_FREE)
3031 mdb_txn_reset(MDB_txn *txn)
3036 /* This call is only valid for read-only txns */
3037 if (!(txn->mt_flags & MDB_TXN_RDONLY))
3040 mdb_txn_end(txn, MDB_END_RESET);
3044 mdb_txn_abort(MDB_txn *txn)
3050 mdb_txn_abort(txn->mt_child);
3052 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
3055 /** Save the freelist as of this transaction to the freeDB.
3056 * This changes the freelist. Keep trying until it stabilizes.
3059 mdb_freelist_save(MDB_txn *txn)
3061 /* env->me_pghead[] can grow and shrink during this call.
3062 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3063 * Page numbers cannot disappear from txn->mt_free_pgs[].
3066 MDB_env *env = txn->mt_env;
3067 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3068 txnid_t pglast = 0, head_id = 0;
3069 pgno_t freecnt = 0, *free_pgs, *mop;
3070 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3072 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3074 if (env->me_pghead) {
3075 /* Make sure first page of freeDB is touched and on freelist */
3076 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3077 if (rc && rc != MDB_NOTFOUND)
3081 if (!env->me_pghead && txn->mt_loose_pgs) {
3082 /* Put loose page numbers in mt_free_pgs, since
3083 * we may be unable to return them to me_pghead.
3085 MDB_page *mp = txn->mt_loose_pgs;
3086 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3088 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3089 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3090 txn->mt_loose_pgs = NULL;
3091 txn->mt_loose_count = 0;
3094 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3095 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3096 ? SSIZE_MAX : maxfree_1pg;
3099 /* Come back here after each Put() in case freelist changed */
3104 /* If using records from freeDB which we have not yet
3105 * deleted, delete them and any we reserved for me_pghead.
3107 while (pglast < env->me_pglast) {
3108 rc = mdb_cursor_first(&mc, &key, NULL);
3111 pglast = head_id = *(txnid_t *)key.mv_data;
3112 total_room = head_room = 0;
3113 mdb_tassert(txn, pglast <= env->me_pglast);
3114 rc = mdb_cursor_del(&mc, 0);
3119 /* Save the IDL of pages freed by this txn, to a single record */
3120 if (freecnt < txn->mt_free_pgs[0]) {
3122 /* Make sure last page of freeDB is touched and on freelist */
3123 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3124 if (rc && rc != MDB_NOTFOUND)
3127 free_pgs = txn->mt_free_pgs;
3128 /* Write to last page of freeDB */
3129 key.mv_size = sizeof(txn->mt_txnid);
3130 key.mv_data = &txn->mt_txnid;
3132 freecnt = free_pgs[0];
3133 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3134 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3137 /* Retry if mt_free_pgs[] grew during the Put() */
3138 free_pgs = txn->mt_free_pgs;
3139 } while (freecnt < free_pgs[0]);
3140 mdb_midl_sort(free_pgs);
3141 memcpy(data.mv_data, free_pgs, data.mv_size);
3144 unsigned int i = free_pgs[0];
3145 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3146 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3148 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3154 mop = env->me_pghead;
3155 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3157 /* Reserve records for me_pghead[]. Split it if multi-page,
3158 * to avoid searching freeDB for a page range. Use keys in
3159 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3161 if (total_room >= mop_len) {
3162 if (total_room == mop_len || --more < 0)
3164 } else if (head_room >= maxfree_1pg && head_id > 1) {
3165 /* Keep current record (overflow page), add a new one */
3169 /* (Re)write {key = head_id, IDL length = head_room} */
3170 total_room -= head_room;
3171 head_room = mop_len - total_room;
3172 if (head_room > maxfree_1pg && head_id > 1) {
3173 /* Overflow multi-page for part of me_pghead */
3174 head_room /= head_id; /* amortize page sizes */
3175 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3176 } else if (head_room < 0) {
3177 /* Rare case, not bothering to delete this record */
3180 key.mv_size = sizeof(head_id);
3181 key.mv_data = &head_id;
3182 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3183 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3186 /* IDL is initially empty, zero out at least the length */
3187 pgs = (pgno_t *)data.mv_data;
3188 j = head_room > clean_limit ? head_room : 0;
3192 total_room += head_room;
3195 /* Return loose page numbers to me_pghead, though usually none are
3196 * left at this point. The pages themselves remain in dirty_list.
3198 if (txn->mt_loose_pgs) {
3199 MDB_page *mp = txn->mt_loose_pgs;
3200 unsigned count = txn->mt_loose_count;
3202 /* Room for loose pages + temp IDL with same */
3203 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3205 mop = env->me_pghead;
3206 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3207 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3208 loose[ ++count ] = mp->mp_pgno;
3210 mdb_midl_sort(loose);
3211 mdb_midl_xmerge(mop, loose);
3212 txn->mt_loose_pgs = NULL;
3213 txn->mt_loose_count = 0;
3217 /* Fill in the reserved me_pghead records */
3223 rc = mdb_cursor_first(&mc, &key, &data);
3224 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3225 txnid_t id = *(txnid_t *)key.mv_data;
3226 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3229 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3231 if (len > mop_len) {
3233 data.mv_size = (len + 1) * sizeof(MDB_ID);
3235 data.mv_data = mop -= len;
3238 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3240 if (rc || !(mop_len -= len))
3247 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3248 * @param[in] txn the transaction that's being committed
3249 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3250 * @return 0 on success, non-zero on failure.
3253 mdb_page_flush(MDB_txn *txn, int keep)
3255 MDB_env *env = txn->mt_env;
3256 MDB_ID2L dl = txn->mt_u.dirty_list;
3257 unsigned psize = env->me_psize, j;
3258 int i, pagecount = dl[0].mid, rc;
3259 size_t size = 0, pos = 0;
3261 MDB_page *dp = NULL;
3265 struct iovec iov[MDB_COMMIT_PAGES];
3266 ssize_t wpos = 0, wsize = 0, wres;
3267 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3273 if (env->me_flags & MDB_WRITEMAP) {
3274 /* Clear dirty flags */
3275 while (++i <= pagecount) {
3277 /* Don't flush this page yet */
3278 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3279 dp->mp_flags &= ~P_KEEP;
3283 dp->mp_flags &= ~P_DIRTY;
3288 /* Write the pages */
3290 if (++i <= pagecount) {
3292 /* Don't flush this page yet */
3293 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3294 dp->mp_flags &= ~P_KEEP;
3299 /* clear dirty flag */
3300 dp->mp_flags &= ~P_DIRTY;
3303 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3308 /* Windows actually supports scatter/gather I/O, but only on
3309 * unbuffered file handles. Since we're relying on the OS page
3310 * cache for all our data, that's self-defeating. So we just
3311 * write pages one at a time. We use the ov structure to set
3312 * the write offset, to at least save the overhead of a Seek
3315 DPRINTF(("committing page %"Z"u", pgno));
3316 memset(&ov, 0, sizeof(ov));
3317 ov.Offset = pos & 0xffffffff;
3318 ov.OffsetHigh = pos >> 16 >> 16;
3319 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3321 DPRINTF(("WriteFile: %d", rc));
3325 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3326 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3329 /* Write previous page(s) */
3330 #ifdef MDB_USE_PWRITEV
3331 wres = pwritev(env->me_fd, iov, n, wpos);
3334 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3337 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3341 DPRINTF(("lseek: %s", strerror(rc)));
3344 wres = writev(env->me_fd, iov, n);
3347 if (wres != wsize) {
3352 DPRINTF(("Write error: %s", strerror(rc)));
3354 rc = EIO; /* TODO: Use which error code? */
3355 DPUTS("short write, filesystem full?");
3366 DPRINTF(("committing page %"Z"u", pgno));
3367 next_pos = pos + size;
3368 iov[n].iov_len = size;
3369 iov[n].iov_base = (char *)dp;
3375 /* MIPS has cache coherency issues, this is a no-op everywhere else
3376 * Note: for any size >= on-chip cache size, entire on-chip cache is
3379 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3381 for (i = keep; ++i <= pagecount; ) {
3383 /* This is a page we skipped above */
3386 dl[j].mid = dp->mp_pgno;
3389 mdb_dpage_free(env, dp);
3394 txn->mt_dirty_room += i - j;
3400 mdb_txn_commit(MDB_txn *txn)
3403 unsigned int i, end_mode;
3409 /* mdb_txn_end() mode for a commit which writes nothing */
3410 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3412 if (txn->mt_child) {
3413 rc = mdb_txn_commit(txn->mt_child);
3420 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3424 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3425 DPUTS("txn has failed/finished, can't commit");
3427 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3432 if (txn->mt_parent) {
3433 MDB_txn *parent = txn->mt_parent;
3437 unsigned x, y, len, ps_len;
3439 /* Append our free list to parent's */
3440 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3443 mdb_midl_free(txn->mt_free_pgs);
3444 /* Failures after this must either undo the changes
3445 * to the parent or set MDB_TXN_ERROR in the parent.
3448 parent->mt_next_pgno = txn->mt_next_pgno;
3449 parent->mt_flags = txn->mt_flags;
3451 /* Merge our cursors into parent's and close them */
3452 mdb_cursors_close(txn, 1);
3454 /* Update parent's DB table. */
3455 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3456 parent->mt_numdbs = txn->mt_numdbs;
3457 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI];
3458 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI];
3459 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3460 /* preserve parent's DB_NEW status */
3461 x = parent->mt_dbflags[i] & DB_NEW;
3462 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3465 dst = parent->mt_u.dirty_list;
3466 src = txn->mt_u.dirty_list;
3467 /* Remove anything in our dirty list from parent's spill list */
3468 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3470 pspill[0] = (pgno_t)-1;
3471 /* Mark our dirty pages as deleted in parent spill list */
3472 for (i=0, len=src[0].mid; ++i <= len; ) {
3473 MDB_ID pn = src[i].mid << 1;
3474 while (pn > pspill[x])
3476 if (pn == pspill[x]) {
3481 /* Squash deleted pagenums if we deleted any */
3482 for (x=y; ++x <= ps_len; )
3483 if (!(pspill[x] & 1))
3484 pspill[++y] = pspill[x];
3488 /* Remove anything in our spill list from parent's dirty list */
3489 if (txn->mt_spill_pgs && txn->mt_spill_pgs[0]) {
3490 for (i=1; i<=txn->mt_spill_pgs[0]; i++) {
3491 MDB_ID pn = txn->mt_spill_pgs[i];
3493 continue; /* deleted spillpg */
3495 y = mdb_mid2l_search(dst, pn);
3496 if (y <= dst[0].mid && dst[y].mid == pn) {
3498 while (y < dst[0].mid) {
3507 /* Find len = length of merging our dirty list with parent's */
3509 dst[0].mid = 0; /* simplify loops */
3510 if (parent->mt_parent) {
3511 len = x + src[0].mid;
3512 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3513 for (i = x; y && i; y--) {
3514 pgno_t yp = src[y].mid;
3515 while (yp < dst[i].mid)
3517 if (yp == dst[i].mid) {
3522 } else { /* Simplify the above for single-ancestor case */
3523 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3525 /* Merge our dirty list with parent's */
3527 for (i = len; y; dst[i--] = src[y--]) {
3528 pgno_t yp = src[y].mid;
3529 while (yp < dst[x].mid)
3530 dst[i--] = dst[x--];
3531 if (yp == dst[x].mid)
3532 free(dst[x--].mptr);
3534 mdb_tassert(txn, i == x);
3536 free(txn->mt_u.dirty_list);
3537 parent->mt_dirty_room = txn->mt_dirty_room;
3538 if (txn->mt_spill_pgs) {
3539 if (parent->mt_spill_pgs) {
3540 /* TODO: Prevent failure here, so parent does not fail */
3541 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3543 parent->mt_flags |= MDB_TXN_ERROR;
3544 mdb_midl_free(txn->mt_spill_pgs);
3545 mdb_midl_sort(parent->mt_spill_pgs);
3547 parent->mt_spill_pgs = txn->mt_spill_pgs;
3551 /* Append our loose page list to parent's */
3552 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(*lp))
3554 *lp = txn->mt_loose_pgs;
3555 parent->mt_loose_count += txn->mt_loose_count;
3557 parent->mt_child = NULL;
3558 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3563 if (txn != env->me_txn) {
3564 DPUTS("attempt to commit unknown transaction");
3569 mdb_cursors_close(txn, 0);
3571 if (!txn->mt_u.dirty_list[0].mid &&
3572 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3575 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3576 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3578 /* Update DB root pointers */
3579 if (txn->mt_numdbs > CORE_DBS) {
3583 data.mv_size = sizeof(MDB_db);
3585 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3586 for (i = CORE_DBS; i < txn->mt_numdbs; i++) {
3587 if (txn->mt_dbflags[i] & DB_DIRTY) {
3588 if (TXN_DBI_CHANGED(txn, i)) {
3592 data.mv_data = &txn->mt_dbs[i];
3593 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3601 rc = mdb_freelist_save(txn);
3605 mdb_midl_free(env->me_pghead);
3606 env->me_pghead = NULL;
3607 mdb_midl_shrink(&txn->mt_free_pgs);
3613 if ((rc = mdb_page_flush(txn, 0)) ||
3614 (rc = mdb_env_sync(env, 0)) ||
3615 (rc = mdb_env_write_meta(txn)))
3617 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3620 mdb_txn_end(txn, end_mode);
3628 /** Read the environment parameters of a DB environment before
3629 * mapping it into memory.
3630 * @param[in] env the environment handle
3631 * @param[out] meta address of where to store the meta information
3632 * @return 0 on success, non-zero on failure.
3635 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3641 enum { Size = sizeof(pbuf) };
3643 /* We don't know the page size yet, so use a minimum value.
3644 * Read both meta pages so we can use the latest one.
3647 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) {
3651 memset(&ov, 0, sizeof(ov));
3653 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3654 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3657 rc = pread(env->me_fd, &pbuf, Size, off);
3660 if (rc == 0 && off == 0)
3662 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3663 DPRINTF(("read: %s", mdb_strerror(rc)));
3667 p = (MDB_page *)&pbuf;
3669 if (!F_ISSET(p->mp_flags, P_META)) {
3670 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3675 if (m->mm_magic != MDB_MAGIC) {
3676 DPUTS("meta has invalid magic");
3680 if (m->mm_version != MDB_DATA_VERSION) {
3681 DPRINTF(("database is version %u, expected version %u",
3682 m->mm_version, MDB_DATA_VERSION));
3683 return MDB_VERSION_MISMATCH;
3686 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3692 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3694 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3696 meta->mm_magic = MDB_MAGIC;
3697 meta->mm_version = MDB_DATA_VERSION;
3698 meta->mm_mapsize = env->me_mapsize;
3699 meta->mm_psize = env->me_psize;
3700 meta->mm_last_pg = NUM_METAS-1;
3701 meta->mm_flags = env->me_flags & 0xffff;
3702 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */
3703 meta->mm_dbs[FREE_DBI].md_root = P_INVALID;
3704 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID;
3707 /** Write the environment parameters of a freshly created DB environment.
3708 * @param[in] env the environment handle
3709 * @param[in] meta the #MDB_meta to write
3710 * @return 0 on success, non-zero on failure.
3713 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3721 memset(&ov, 0, sizeof(ov));
3722 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3724 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3727 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3728 len = pwrite(fd, ptr, size, pos); \
3729 if (len == -1 && ErrCode() == EINTR) continue; \
3730 rc = (len >= 0); break; } while(1)
3733 DPUTS("writing new meta page");
3735 psize = env->me_psize;
3737 p = calloc(NUM_METAS, psize);
3742 p->mp_flags = P_META;
3743 *(MDB_meta *)METADATA(p) = *meta;
3745 q = (MDB_page *)((char *)p + psize);
3747 q->mp_flags = P_META;
3748 *(MDB_meta *)METADATA(q) = *meta;
3750 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0);
3753 else if ((unsigned) len == psize * NUM_METAS)
3761 /** Update the environment info to commit a transaction.
3762 * @param[in] txn the transaction that's being committed
3763 * @return 0 on success, non-zero on failure.
3766 mdb_env_write_meta(MDB_txn *txn)
3769 MDB_meta meta, metab, *mp;
3773 int rc, len, toggle;
3782 toggle = txn->mt_txnid & 1;
3783 DPRINTF(("writing meta page %d for root page %"Z"u",
3784 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3787 flags = env->me_flags;
3788 mp = env->me_metas[toggle];
3789 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3790 /* Persist any increases of mapsize config */
3791 if (mapsize < env->me_mapsize)
3792 mapsize = env->me_mapsize;
3794 if (flags & MDB_WRITEMAP) {
3795 mp->mm_mapsize = mapsize;
3796 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3797 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3798 mp->mm_last_pg = txn->mt_next_pgno - 1;
3799 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3800 !(defined(__i386__) || defined(__x86_64__))
3801 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3802 __sync_synchronize();
3804 mp->mm_txnid = txn->mt_txnid;
3805 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3806 unsigned meta_size = env->me_psize;
3807 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3808 ptr = (char *)mp - PAGEHDRSZ;
3809 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3810 r2 = (ptr - env->me_map) & (env->me_os_psize - 1);
3814 if (MDB_MSYNC(ptr, meta_size, rc)) {
3821 metab.mm_txnid = mp->mm_txnid;
3822 metab.mm_last_pg = mp->mm_last_pg;
3824 meta.mm_mapsize = mapsize;
3825 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3826 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3827 meta.mm_last_pg = txn->mt_next_pgno - 1;
3828 meta.mm_txnid = txn->mt_txnid;
3830 off = offsetof(MDB_meta, mm_mapsize);
3831 ptr = (char *)&meta + off;
3832 len = sizeof(MDB_meta) - off;
3833 off += (char *)mp - env->me_map;
3835 /* Write to the SYNC fd unless MDB_NOSYNC/MDB_NOMETASYNC.
3836 * (me_mfd goes to the same file as me_fd, but writing to it
3837 * also syncs to disk. Avoids a separate fdatasync() call.)
3839 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3842 memset(&ov, 0, sizeof(ov));
3844 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3849 rc = pwrite(mfd, ptr, len, off);
3852 rc = rc < 0 ? ErrCode() : EIO;
3857 DPUTS("write failed, disk error?");
3858 /* On a failure, the pagecache still contains the new data.
3859 * Write some old data back, to prevent it from being used.
3860 * Use the non-SYNC fd; we know it will fail anyway.
3862 meta.mm_last_pg = metab.mm_last_pg;
3863 meta.mm_txnid = metab.mm_txnid;
3865 memset(&ov, 0, sizeof(ov));
3867 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3869 r2 = pwrite(env->me_fd, ptr, len, off);
3870 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3873 env->me_flags |= MDB_FATAL_ERROR;
3876 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3877 CACHEFLUSH(env->me_map + off, len, DCACHE);
3879 /* Memory ordering issues are irrelevant; since the entire writer
3880 * is wrapped by wmutex, all of these changes will become visible
3881 * after the wmutex is unlocked. Since the DB is multi-version,
3882 * readers will get consistent data regardless of how fresh or
3883 * how stale their view of these values is.
3886 env->me_txns->mti_txnid = txn->mt_txnid;
3891 /** Check both meta pages to see which one is newer.
3892 * @param[in] env the environment handle
3893 * @return newest #MDB_meta.
3896 mdb_env_pick_meta(const MDB_env *env)
3898 MDB_meta *const *metas = env->me_metas;
3899 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
3903 mdb_env_create(MDB_env **env)
3907 e = calloc(1, sizeof(MDB_env));
3911 e->me_maxreaders = DEFAULT_READERS;
3912 e->me_maxdbs = e->me_numdbs = CORE_DBS;
3913 e->me_fd = INVALID_HANDLE_VALUE;
3914 e->me_lfd = INVALID_HANDLE_VALUE;
3915 e->me_mfd = INVALID_HANDLE_VALUE;
3916 #ifdef MDB_USE_POSIX_SEM
3917 e->me_rmutex = SEM_FAILED;
3918 e->me_wmutex = SEM_FAILED;
3920 e->me_pid = getpid();
3921 GET_PAGESIZE(e->me_os_psize);
3922 VGMEMP_CREATE(e,0,0);
3928 mdb_env_map(MDB_env *env, void *addr)
3931 unsigned int flags = env->me_flags;
3935 LONG sizelo, sizehi;
3938 if (flags & MDB_RDONLY) {
3939 /* Don't set explicit map size, use whatever exists */
3944 msize = env->me_mapsize;
3945 sizelo = msize & 0xffffffff;
3946 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3948 /* Windows won't create mappings for zero length files.
3949 * and won't map more than the file size.
3950 * Just set the maxsize right now.
3952 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3953 || !SetEndOfFile(env->me_fd)
3954 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3958 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3959 PAGE_READWRITE : PAGE_READONLY,
3960 sizehi, sizelo, NULL);
3963 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3964 FILE_MAP_WRITE : FILE_MAP_READ,
3966 rc = env->me_map ? 0 : ErrCode();
3971 int prot = PROT_READ;
3972 if (flags & MDB_WRITEMAP) {
3974 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3977 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3979 if (env->me_map == MAP_FAILED) {
3984 if (flags & MDB_NORDAHEAD) {
3985 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3987 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3989 #ifdef POSIX_MADV_RANDOM
3990 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3991 #endif /* POSIX_MADV_RANDOM */
3992 #endif /* MADV_RANDOM */
3996 /* Can happen because the address argument to mmap() is just a
3997 * hint. mmap() can pick another, e.g. if the range is in use.
3998 * The MAP_FIXED flag would prevent that, but then mmap could
3999 * instead unmap existing pages to make room for the new map.
4001 if (addr && env->me_map != addr)
4002 return EBUSY; /* TODO: Make a new MDB_* error code? */
4004 p = (MDB_page *)env->me_map;
4005 env->me_metas[0] = METADATA(p);
4006 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
4012 mdb_env_set_mapsize(MDB_env *env, size_t size)
4014 /* If env is already open, caller is responsible for making
4015 * sure there are no active txns.
4023 meta = mdb_env_pick_meta(env);
4025 size = meta->mm_mapsize;
4027 /* Silently round up to minimum if the size is too small */
4028 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
4032 munmap(env->me_map, env->me_mapsize);
4033 env->me_mapsize = size;
4034 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
4035 rc = mdb_env_map(env, old);
4039 env->me_mapsize = size;
4041 env->me_maxpg = env->me_mapsize / env->me_psize;
4046 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
4050 env->me_maxdbs = dbs + CORE_DBS;
4055 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
4057 if (env->me_map || readers < 1)
4059 env->me_maxreaders = readers;
4064 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4066 if (!env || !readers)
4068 *readers = env->me_maxreaders;
4073 mdb_fsize(HANDLE fd, size_t *size)
4076 LARGE_INTEGER fsize;
4078 if (!GetFileSizeEx(fd, &fsize))
4081 *size = fsize.QuadPart;
4095 typedef wchar_t mdb_nchar_t;
4096 # define MDB_NAME(str) L##str
4097 # define mdb_name_cpy wcscpy
4099 /** Character type for file names: char on Unix, wchar_t on Windows */
4100 typedef char mdb_nchar_t;
4101 # define MDB_NAME(str) str /**< #mdb_nchar_t[] string literal */
4102 # define mdb_name_cpy strcpy /**< Copy name (#mdb_nchar_t string) */
4105 /** Filename - string of #mdb_nchar_t[] */
4106 typedef struct MDB_name {
4107 int mn_len; /**< Length */
4108 int mn_alloced; /**< True if #mn_val was malloced */
4109 mdb_nchar_t *mn_val; /**< Contents */
4112 /** Filename suffixes [datafile,lockfile][without,with MDB_NOSUBDIR] */
4113 static const mdb_nchar_t *const mdb_suffixes[2][2] = {
4114 { MDB_NAME("/data.mdb"), MDB_NAME("") },
4115 { MDB_NAME("/lock.mdb"), MDB_NAME("-lock") }
4118 #define MDB_SUFFLEN 9 /**< Max string length in #mdb_suffixes[] */
4120 /** Set up filename + scratch area for filename suffix, for opening files.
4121 * It should be freed with #mdb_fname_destroy().
4122 * On Windows, paths are converted from char *UTF-8 to wchar_t *UTF-16.
4124 * @param[in] path Pathname for #mdb_env_open().
4125 * @param[in] envflags Whether a subdir and/or lockfile will be used.
4126 * @param[out] fname Resulting filename, with room for a suffix if necessary.
4129 mdb_fname_init(const char *path, unsigned envflags, MDB_name *fname)
4131 int no_suffix = F_ISSET(envflags, MDB_NOSUBDIR|MDB_NOLOCK);
4132 fname->mn_alloced = 0;
4134 return utf8_to_utf16(path, fname, no_suffix ? 0 : MDB_SUFFLEN);
4136 fname->mn_len = strlen(path);
4138 fname->mn_val = (char *) path;
4139 else if ((fname->mn_val = malloc(fname->mn_len + MDB_SUFFLEN+1)) != NULL) {
4140 fname->mn_alloced = 1;
4141 strcpy(fname->mn_val, path);
4149 /** Destroy \b fname from #mdb_fname_init() */
4150 #define mdb_fname_destroy(fname) \
4151 do { if ((fname).mn_alloced) free((fname).mn_val); } while (0)
4153 #ifdef O_CLOEXEC /* POSIX.1-2008: Set FD_CLOEXEC atomically at open() */
4154 # define MDB_CLOEXEC O_CLOEXEC
4156 # define MDB_CLOEXEC 0
4159 /** File type, access mode etc. for #mdb_fopen() */
4160 enum mdb_fopen_type {
4162 MDB_O_RDONLY, MDB_O_RDWR, MDB_O_META, MDB_O_COPY, MDB_O_LOCKS
4164 /* A comment in mdb_fopen() explains some O_* flag choices. */
4165 MDB_O_RDONLY= O_RDONLY, /**< for RDONLY me_fd */
4166 MDB_O_RDWR = O_RDWR |O_CREAT, /**< for me_fd */
4167 MDB_O_META = O_WRONLY|MDB_DSYNC |MDB_CLOEXEC, /**< for me_mfd */
4168 MDB_O_COPY = O_WRONLY|O_CREAT|O_EXCL|MDB_CLOEXEC, /**< for #mdb_env_copy() */
4169 /** Bitmask for open() flags in enum #mdb_fopen_type. The other bits
4170 * distinguish otherwise-equal MDB_O_* constants from each other.
4172 MDB_O_MASK = MDB_O_RDWR|MDB_CLOEXEC | MDB_O_RDONLY|MDB_O_META|MDB_O_COPY,
4173 MDB_O_LOCKS = MDB_O_RDWR|MDB_CLOEXEC | ((MDB_O_MASK+1) & ~MDB_O_MASK) /**< for me_lfd */
4177 /** Open an LMDB file.
4178 * @param[in] env The LMDB environment.
4179 * @param[in,out] fname Path from from #mdb_fname_init(). A suffix is
4180 * appended if necessary to create the filename, without changing mn_len.
4181 * @param[in] which Determines file type, access mode, etc.
4182 * @param[in] mode The Unix permissions for the file, if we create it.
4183 * @param[out] res Resulting file handle.
4184 * @return 0 on success, non-zero on failure.
4187 mdb_fopen(const MDB_env *env, MDB_name *fname,
4188 enum mdb_fopen_type which, mdb_mode_t mode,
4191 int rc = MDB_SUCCESS;
4194 DWORD acc, share, disp, attrs;
4199 if (fname->mn_alloced) /* modifiable copy */
4200 mdb_name_cpy(fname->mn_val + fname->mn_len,
4201 mdb_suffixes[which==MDB_O_LOCKS][F_ISSET(env->me_flags, MDB_NOSUBDIR)]);
4203 /* The directory must already exist. Usually the file need not.
4204 * MDB_O_META requires the file because we already created it using
4205 * MDB_O_RDWR. MDB_O_COPY must not overwrite an existing file.
4207 * With MDB_O_COPY we do not want the OS to cache the writes, since
4208 * the source data is already in the OS cache.
4210 * The lockfile needs FD_CLOEXEC (close file descriptor on exec*())
4211 * to avoid the flock() issues noted under Caveats in lmdb.h.
4212 * Also set it for other filehandles which the user cannot get at
4213 * and close himself, which he may need after fork(). I.e. all but
4214 * me_fd, which programs do use via mdb_env_get_fd().
4218 acc = GENERIC_READ|GENERIC_WRITE;
4219 share = FILE_SHARE_READ|FILE_SHARE_WRITE;
4221 attrs = FILE_ATTRIBUTE_NORMAL;
4223 case MDB_O_RDONLY: /* read-only datafile */
4225 disp = OPEN_EXISTING;
4227 case MDB_O_META: /* for writing metapages */
4228 acc = GENERIC_WRITE;
4229 disp = OPEN_EXISTING;
4230 attrs = FILE_ATTRIBUTE_NORMAL|FILE_FLAG_WRITE_THROUGH;
4232 case MDB_O_COPY: /* mdb_env_copy() & co */
4233 acc = GENERIC_WRITE;
4236 attrs = FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH;
4238 default: break; /* silence gcc -Wswitch (not all enum values handled) */
4240 fd = CreateFileW(fname->mn_val, acc, share, NULL, disp, attrs, NULL);
4242 fd = open(fname->mn_val, which & MDB_O_MASK, mode);
4245 if (fd == INVALID_HANDLE_VALUE)
4249 if (which != MDB_O_RDONLY && which != MDB_O_RDWR) {
4250 /* Set CLOEXEC if we could not pass it to open() */
4251 if (!MDB_CLOEXEC && (flags = fcntl(fd, F_GETFD)) != -1)
4252 (void) fcntl(fd, F_SETFD, flags | FD_CLOEXEC);
4254 if (which == MDB_O_COPY && env->me_psize >= env->me_os_psize) {
4255 /* This may require buffer alignment. There is no portable
4256 * way to ask how much, so we require OS pagesize alignment.
4258 # ifdef F_NOCACHE /* __APPLE__ */
4259 (void) fcntl(fd, F_NOCACHE, 1);
4260 # elif defined O_DIRECT
4261 /* open(...O_DIRECT...) would break on filesystems without
4262 * O_DIRECT support (ITS#7682). Try to set it here instead.
4264 if ((flags = fcntl(fd, F_GETFL)) != -1)
4265 (void) fcntl(fd, F_SETFL, flags | O_DIRECT);
4269 #endif /* !_WIN32 */
4276 #ifdef BROKEN_FDATASYNC
4277 #include <sys/utsname.h>
4278 #include <sys/vfs.h>
4281 /** Further setup required for opening an LMDB environment
4284 mdb_env_open2(MDB_env *env)
4286 unsigned int flags = env->me_flags;
4287 int i, newenv = 0, rc;
4291 /* See if we should use QueryLimited */
4293 if ((rc & 0xff) > 5)
4294 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4296 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4299 #ifdef BROKEN_FDATASYNC
4300 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4301 * https://lkml.org/lkml/2012/9/3/83
4302 * Kernels after 3.6-rc6 are known good.
4303 * https://lkml.org/lkml/2012/9/10/556
4304 * See if the DB is on ext3/ext4, then check for new enough kernel
4305 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4310 fstatfs(env->me_fd, &st);
4311 while (st.f_type == 0xEF53) {
4315 if (uts.release[0] < '3') {
4316 if (!strncmp(uts.release, "2.6.32.", 7)) {
4317 i = atoi(uts.release+7);
4319 break; /* 2.6.32.60 and newer is OK */
4320 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4321 i = atoi(uts.release+7);
4323 break; /* 2.6.34.15 and newer is OK */
4325 } else if (uts.release[0] == '3') {
4326 i = atoi(uts.release+2);
4328 break; /* 3.6 and newer is OK */
4330 i = atoi(uts.release+4);
4332 break; /* 3.5.4 and newer is OK */
4333 } else if (i == 2) {
4334 i = atoi(uts.release+4);
4336 break; /* 3.2.30 and newer is OK */
4338 } else { /* 4.x and newer is OK */
4341 env->me_flags |= MDB_FSYNCONLY;
4347 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4350 DPUTS("new mdbenv");
4352 env->me_psize = env->me_os_psize;
4353 if (env->me_psize > MAX_PAGESIZE)
4354 env->me_psize = MAX_PAGESIZE;
4355 memset(&meta, 0, sizeof(meta));
4356 mdb_env_init_meta0(env, &meta);
4357 meta.mm_mapsize = DEFAULT_MAPSIZE;
4359 env->me_psize = meta.mm_psize;
4362 /* Was a mapsize configured? */
4363 if (!env->me_mapsize) {
4364 env->me_mapsize = meta.mm_mapsize;
4367 /* Make sure mapsize >= committed data size. Even when using
4368 * mm_mapsize, which could be broken in old files (ITS#7789).
4370 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4371 if (env->me_mapsize < minsize)
4372 env->me_mapsize = minsize;
4374 meta.mm_mapsize = env->me_mapsize;
4376 if (newenv && !(flags & MDB_FIXEDMAP)) {
4377 /* mdb_env_map() may grow the datafile. Write the metapages
4378 * first, so the file will be valid if initialization fails.
4379 * Except with FIXEDMAP, since we do not yet know mm_address.
4380 * We could fill in mm_address later, but then a different
4381 * program might end up doing that - one with a memory layout
4382 * and map address which does not suit the main program.
4384 rc = mdb_env_init_meta(env, &meta);
4390 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4395 if (flags & MDB_FIXEDMAP)
4396 meta.mm_address = env->me_map;
4397 i = mdb_env_init_meta(env, &meta);
4398 if (i != MDB_SUCCESS) {
4403 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4404 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4406 #if !(MDB_MAXKEYSIZE)
4407 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4409 env->me_maxpg = env->me_mapsize / env->me_psize;
4413 MDB_meta *meta = mdb_env_pick_meta(env);
4414 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4416 DPRINTF(("opened database version %u, pagesize %u",
4417 meta->mm_version, env->me_psize));
4418 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4419 DPRINTF(("depth: %u", db->md_depth));
4420 DPRINTF(("entries: %"Z"u", db->md_entries));
4421 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4422 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4423 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4424 DPRINTF(("root: %"Z"u", db->md_root));
4432 /** Release a reader thread's slot in the reader lock table.
4433 * This function is called automatically when a thread exits.
4434 * @param[in] ptr This points to the slot in the reader lock table.
4437 mdb_env_reader_dest(void *ptr)
4439 MDB_reader *reader = ptr;
4442 if (reader->mr_pid == getpid()) /* catch pthread_exit() in child process */
4444 /* We omit the mutex, so do this atomically (i.e. skip mr_txnid) */
4449 /** Junk for arranging thread-specific callbacks on Windows. This is
4450 * necessarily platform and compiler-specific. Windows supports up
4451 * to 1088 keys. Let's assume nobody opens more than 64 environments
4452 * in a single process, for now. They can override this if needed.
4454 #ifndef MAX_TLS_KEYS
4455 #define MAX_TLS_KEYS 64
4457 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4458 static int mdb_tls_nkeys;
4460 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4464 case DLL_PROCESS_ATTACH: break;
4465 case DLL_THREAD_ATTACH: break;
4466 case DLL_THREAD_DETACH:
4467 for (i=0; i<mdb_tls_nkeys; i++) {
4468 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4470 mdb_env_reader_dest(r);
4474 case DLL_PROCESS_DETACH: break;
4479 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4481 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4485 /* Force some symbol references.
4486 * _tls_used forces the linker to create the TLS directory if not already done
4487 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4489 #pragma comment(linker, "/INCLUDE:_tls_used")
4490 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4491 #pragma const_seg(".CRT$XLB")
4492 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4493 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4496 #pragma comment(linker, "/INCLUDE:__tls_used")
4497 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4498 #pragma data_seg(".CRT$XLB")
4499 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4501 #endif /* WIN 32/64 */
4502 #endif /* !__GNUC__ */
4505 /** Downgrade the exclusive lock on the region back to shared */
4507 mdb_env_share_locks(MDB_env *env, int *excl)
4510 MDB_meta *meta = mdb_env_pick_meta(env);
4512 env->me_txns->mti_txnid = meta->mm_txnid;
4517 /* First acquire a shared lock. The Unlock will
4518 * then release the existing exclusive lock.
4520 memset(&ov, 0, sizeof(ov));
4521 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4524 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4530 struct flock lock_info;
4531 /* The shared lock replaces the existing lock */
4532 memset((void *)&lock_info, 0, sizeof(lock_info));
4533 lock_info.l_type = F_RDLCK;
4534 lock_info.l_whence = SEEK_SET;
4535 lock_info.l_start = 0;
4536 lock_info.l_len = 1;
4537 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4538 (rc = ErrCode()) == EINTR) ;
4539 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4546 /** Try to get exclusive lock, otherwise shared.
4547 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4550 mdb_env_excl_lock(MDB_env *env, int *excl)
4554 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4558 memset(&ov, 0, sizeof(ov));
4559 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4566 struct flock lock_info;
4567 memset((void *)&lock_info, 0, sizeof(lock_info));
4568 lock_info.l_type = F_WRLCK;
4569 lock_info.l_whence = SEEK_SET;
4570 lock_info.l_start = 0;
4571 lock_info.l_len = 1;
4572 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4573 (rc = ErrCode()) == EINTR) ;
4577 # ifndef MDB_USE_POSIX_MUTEX
4578 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4581 lock_info.l_type = F_RDLCK;
4582 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4583 (rc = ErrCode()) == EINTR) ;
4593 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4595 * @(#) $Revision: 5.1 $
4596 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4597 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4599 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4603 * Please do not copyright this code. This code is in the public domain.
4605 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4606 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4607 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4608 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4609 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4610 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4611 * PERFORMANCE OF THIS SOFTWARE.
4614 * chongo <Landon Curt Noll> /\oo/\
4615 * http://www.isthe.com/chongo/
4617 * Share and Enjoy! :-)
4620 typedef unsigned long long mdb_hash_t;
4621 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4623 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4624 * @param[in] val value to hash
4625 * @param[in] hval initial value for hash
4626 * @return 64 bit hash
4628 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4629 * hval arg on the first call.
4632 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4634 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4635 unsigned char *end = s + val->mv_size;
4637 * FNV-1a hash each octet of the string
4640 /* xor the bottom with the current octet */
4641 hval ^= (mdb_hash_t)*s++;
4643 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4644 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4645 (hval << 7) + (hval << 8) + (hval << 40);
4647 /* return our new hash value */
4651 /** Hash the string and output the encoded hash.
4652 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4653 * very short name limits. We don't care about the encoding being reversible,
4654 * we just want to preserve as many bits of the input as possible in a
4655 * small printable string.
4656 * @param[in] str string to hash
4657 * @param[out] encbuf an array of 11 chars to hold the hash
4659 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4662 mdb_pack85(unsigned long l, char *out)
4666 for (i=0; i<5; i++) {
4667 *out++ = mdb_a85[l % 85];
4673 mdb_hash_enc(MDB_val *val, char *encbuf)
4675 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4677 mdb_pack85(h, encbuf);
4678 mdb_pack85(h>>32, encbuf+5);
4683 /** Open and/or initialize the lock region for the environment.
4684 * @param[in] env The LMDB environment.
4685 * @param[in] fname Filename + scratch area, from #mdb_fname_init().
4686 * @param[in] mode The Unix permissions for the file, if we create it.
4687 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4688 * @return 0 on success, non-zero on failure.
4691 mdb_env_setup_locks(MDB_env *env, MDB_name *fname, int mode, int *excl)
4694 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4696 # define MDB_ERRCODE_ROFS EROFS
4701 rc = mdb_fopen(env, fname, MDB_O_LOCKS, mode, &env->me_lfd);
4703 /* Omit lockfile if read-only env on read-only filesystem */
4704 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4710 if (!(env->me_flags & MDB_NOTLS)) {
4711 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4714 env->me_flags |= MDB_ENV_TXKEY;
4716 /* Windows TLS callbacks need help finding their TLS info. */
4717 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4721 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4725 /* Try to get exclusive lock. If we succeed, then
4726 * nobody is using the lock region and we should initialize it.
4728 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4731 size = GetFileSize(env->me_lfd, NULL);
4733 size = lseek(env->me_lfd, 0, SEEK_END);
4734 if (size == -1) goto fail_errno;
4736 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4737 if (size < rsize && *excl > 0) {
4739 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4740 || !SetEndOfFile(env->me_lfd))
4743 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4747 size = rsize - sizeof(MDB_txninfo);
4748 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4753 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4755 if (!mh) goto fail_errno;
4756 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4758 if (!env->me_txns) goto fail_errno;
4760 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4762 if (m == MAP_FAILED) goto fail_errno;
4768 BY_HANDLE_FILE_INFORMATION stbuf;
4777 if (!mdb_sec_inited) {
4778 InitializeSecurityDescriptor(&mdb_null_sd,
4779 SECURITY_DESCRIPTOR_REVISION);
4780 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4781 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4782 mdb_all_sa.bInheritHandle = FALSE;
4783 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4786 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4787 idbuf.volume = stbuf.dwVolumeSerialNumber;
4788 idbuf.nhigh = stbuf.nFileIndexHigh;
4789 idbuf.nlow = stbuf.nFileIndexLow;
4790 val.mv_data = &idbuf;
4791 val.mv_size = sizeof(idbuf);
4792 mdb_hash_enc(&val, encbuf);
4793 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4794 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4795 env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4796 if (!env->me_rmutex) goto fail_errno;
4797 env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4798 if (!env->me_wmutex) goto fail_errno;
4799 #elif defined(MDB_USE_POSIX_SEM)
4808 #if defined(__NetBSD__)
4809 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4811 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4812 idbuf.dev = stbuf.st_dev;
4813 idbuf.ino = stbuf.st_ino;
4814 val.mv_data = &idbuf;
4815 val.mv_size = sizeof(idbuf);
4816 mdb_hash_enc(&val, encbuf);
4817 #ifdef MDB_SHORT_SEMNAMES
4818 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4820 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4821 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4822 /* Clean up after a previous run, if needed: Try to
4823 * remove both semaphores before doing anything else.
4825 sem_unlink(env->me_txns->mti_rmname);
4826 sem_unlink(env->me_txns->mti_wmname);
4827 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4828 O_CREAT|O_EXCL, mode, 1);
4829 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4830 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4831 O_CREAT|O_EXCL, mode, 1);
4832 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4833 #else /* MDB_USE_POSIX_MUTEX: */
4834 pthread_mutexattr_t mattr;
4836 /* Solaris needs this before initing a robust mutex. Otherwise
4837 * it may skip the init and return EBUSY "seems someone already
4838 * inited" or EINVAL "it was inited differently".
4840 memset(env->me_txns->mti_rmutex, 0, sizeof(*env->me_txns->mti_rmutex));
4841 memset(env->me_txns->mti_wmutex, 0, sizeof(*env->me_txns->mti_wmutex));
4843 if ((rc = pthread_mutexattr_init(&mattr)))
4846 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
4847 #ifdef MDB_ROBUST_SUPPORTED
4848 if (!rc) rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST);
4850 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr);
4851 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr);
4852 pthread_mutexattr_destroy(&mattr);
4855 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4857 env->me_txns->mti_magic = MDB_MAGIC;
4858 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4859 env->me_txns->mti_txnid = 0;
4860 env->me_txns->mti_numreaders = 0;
4863 if (env->me_txns->mti_magic != MDB_MAGIC) {
4864 DPUTS("lock region has invalid magic");
4868 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4869 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4870 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4871 rc = MDB_VERSION_MISMATCH;
4875 if (rc && rc != EACCES && rc != EAGAIN) {
4879 env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4880 if (!env->me_rmutex) goto fail_errno;
4881 env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4882 if (!env->me_wmutex) goto fail_errno;
4883 #elif defined(MDB_USE_POSIX_SEM)
4884 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4885 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4886 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4887 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4898 /** Only a subset of the @ref mdb_env flags can be changed
4899 * at runtime. Changing other flags requires closing the
4900 * environment and re-opening it with the new flags.
4902 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4903 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4904 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4906 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4907 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4911 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4916 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4919 flags |= env->me_flags;
4921 rc = mdb_fname_init(path, flags, &fname);
4925 if (flags & MDB_RDONLY) {
4926 /* silently ignore WRITEMAP when we're only getting read access */
4927 flags &= ~MDB_WRITEMAP;
4929 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4930 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4933 env->me_flags = flags |= MDB_ENV_ACTIVE;
4937 env->me_path = strdup(path);
4938 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4939 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4940 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4941 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4945 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4947 /* For RDONLY, get lockfile after we know datafile exists */
4948 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4949 rc = mdb_env_setup_locks(env, &fname, mode, &excl);
4954 rc = mdb_fopen(env, &fname,
4955 (flags & MDB_RDONLY) ? MDB_O_RDONLY : MDB_O_RDWR,
4960 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4961 rc = mdb_env_setup_locks(env, &fname, mode, &excl);
4966 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4967 if (!(flags & (MDB_RDONLY|MDB_WRITEMAP))) {
4968 /* Synchronous fd for meta writes. Needed even with
4969 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4971 rc = mdb_fopen(env, &fname, MDB_O_META, mode, &env->me_mfd);
4975 DPRINTF(("opened dbenv %p", (void *) env));
4977 rc = mdb_env_share_locks(env, &excl);
4981 if (!(flags & MDB_RDONLY)) {
4983 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4984 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4985 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4986 (txn = calloc(1, size)))
4988 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4989 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4990 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4991 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4993 txn->mt_dbxs = env->me_dbxs;
4994 txn->mt_flags = MDB_TXN_FINISHED;
5004 mdb_env_close0(env, excl);
5006 mdb_fname_destroy(fname);
5010 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
5012 mdb_env_close0(MDB_env *env, int excl)
5016 if (!(env->me_flags & MDB_ENV_ACTIVE))
5019 /* Doing this here since me_dbxs may not exist during mdb_env_close */
5021 for (i = env->me_maxdbs; --i >= CORE_DBS; )
5022 free(env->me_dbxs[i].md_name.mv_data);
5027 free(env->me_dbiseqs);
5028 free(env->me_dbflags);
5030 free(env->me_dirty_list);
5032 mdb_midl_free(env->me_free_pgs);
5034 if (env->me_flags & MDB_ENV_TXKEY) {
5035 pthread_key_delete(env->me_txkey);
5037 /* Delete our key from the global list */
5038 for (i=0; i<mdb_tls_nkeys; i++)
5039 if (mdb_tls_keys[i] == env->me_txkey) {
5040 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
5048 munmap(env->me_map, env->me_mapsize);
5050 if (env->me_mfd != INVALID_HANDLE_VALUE)
5051 (void) close(env->me_mfd);
5052 if (env->me_fd != INVALID_HANDLE_VALUE)
5053 (void) close(env->me_fd);
5055 MDB_PID_T pid = env->me_pid;
5056 /* Clearing readers is done in this function because
5057 * me_txkey with its destructor must be disabled first.
5059 * We skip the the reader mutex, so we touch only
5060 * data owned by this process (me_close_readers and
5061 * our readers), and clear each reader atomically.
5063 for (i = env->me_close_readers; --i >= 0; )
5064 if (env->me_txns->mti_readers[i].mr_pid == pid)
5065 env->me_txns->mti_readers[i].mr_pid = 0;
5067 if (env->me_rmutex) {
5068 CloseHandle(env->me_rmutex);
5069 if (env->me_wmutex) CloseHandle(env->me_wmutex);
5071 /* Windows automatically destroys the mutexes when
5072 * the last handle closes.
5074 #elif defined(MDB_USE_POSIX_SEM)
5075 if (env->me_rmutex != SEM_FAILED) {
5076 sem_close(env->me_rmutex);
5077 if (env->me_wmutex != SEM_FAILED)
5078 sem_close(env->me_wmutex);
5079 /* If we have the filelock: If we are the
5080 * only remaining user, clean up semaphores.
5083 mdb_env_excl_lock(env, &excl);
5085 sem_unlink(env->me_txns->mti_rmname);
5086 sem_unlink(env->me_txns->mti_wmname);
5090 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
5092 if (env->me_lfd != INVALID_HANDLE_VALUE) {
5095 /* Unlock the lockfile. Windows would have unlocked it
5096 * after closing anyway, but not necessarily at once.
5098 UnlockFile(env->me_lfd, 0, 0, 1, 0);
5101 (void) close(env->me_lfd);
5104 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
5108 mdb_env_close(MDB_env *env)
5115 VGMEMP_DESTROY(env);
5116 while ((dp = env->me_dpages) != NULL) {
5117 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
5118 env->me_dpages = dp->mp_next;
5122 mdb_env_close0(env, 0);
5126 /** Compare two items pointing at aligned size_t's */
5128 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
5130 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
5131 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
5134 /** Compare two items pointing at aligned unsigned int's.
5136 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
5137 * but #mdb_cmp_clong() is called instead if the data type is size_t.
5140 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
5142 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
5143 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
5146 /** Compare two items pointing at unsigned ints of unknown alignment.
5147 * Nodes and keys are guaranteed to be 2-byte aligned.
5150 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5152 #if BYTE_ORDER == LITTLE_ENDIAN
5153 unsigned short *u, *c;
5156 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5157 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5160 } while(!x && u > (unsigned short *)a->mv_data);
5163 unsigned short *u, *c, *end;
5166 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5167 u = (unsigned short *)a->mv_data;
5168 c = (unsigned short *)b->mv_data;
5171 } while(!x && u < end);
5176 /** Compare two items lexically */
5178 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5185 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5191 diff = memcmp(a->mv_data, b->mv_data, len);
5192 return diff ? diff : len_diff<0 ? -1 : len_diff;
5195 /** Compare two items in reverse byte order */
5197 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5199 const unsigned char *p1, *p2, *p1_lim;
5203 p1_lim = (const unsigned char *)a->mv_data;
5204 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5205 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5207 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5213 while (p1 > p1_lim) {
5214 diff = *--p1 - *--p2;
5218 return len_diff<0 ? -1 : len_diff;
5221 /** Search for key within a page, using binary search.
5222 * Returns the smallest entry larger or equal to the key.
5223 * If exactp is non-null, stores whether the found entry was an exact match
5224 * in *exactp (1 or 0).
5225 * Updates the cursor index with the index of the found entry.
5226 * If no entry larger or equal to the key is found, returns NULL.
5229 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5231 unsigned int i = 0, nkeys;
5234 MDB_page *mp = mc->mc_pg[mc->mc_top];
5235 MDB_node *node = NULL;
5240 nkeys = NUMKEYS(mp);
5242 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5243 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5246 low = IS_LEAF(mp) ? 0 : 1;
5248 cmp = mc->mc_dbx->md_cmp;
5250 /* Branch pages have no data, so if using integer keys,
5251 * alignment is guaranteed. Use faster mdb_cmp_int.
5253 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5254 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5261 nodekey.mv_size = mc->mc_db->md_pad;
5262 node = NODEPTR(mp, 0); /* fake */
5263 while (low <= high) {
5264 i = (low + high) >> 1;
5265 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5266 rc = cmp(key, &nodekey);
5267 DPRINTF(("found leaf index %u [%s], rc = %i",
5268 i, DKEY(&nodekey), rc));
5277 while (low <= high) {
5278 i = (low + high) >> 1;
5280 node = NODEPTR(mp, i);
5281 nodekey.mv_size = NODEKSZ(node);
5282 nodekey.mv_data = NODEKEY(node);
5284 rc = cmp(key, &nodekey);
5287 DPRINTF(("found leaf index %u [%s], rc = %i",
5288 i, DKEY(&nodekey), rc));
5290 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5291 i, DKEY(&nodekey), NODEPGNO(node), rc));
5302 if (rc > 0) { /* Found entry is less than the key. */
5303 i++; /* Skip to get the smallest entry larger than key. */
5305 node = NODEPTR(mp, i);
5308 *exactp = (rc == 0 && nkeys > 0);
5309 /* store the key index */
5310 mc->mc_ki[mc->mc_top] = i;
5312 /* There is no entry larger or equal to the key. */
5315 /* nodeptr is fake for LEAF2 */
5321 mdb_cursor_adjust(MDB_cursor *mc, func)
5325 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5326 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5333 /** Pop a page off the top of the cursor's stack. */
5335 mdb_cursor_pop(MDB_cursor *mc)
5338 DPRINTF(("popping page %"Z"u off db %d cursor %p",
5339 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5345 mc->mc_flags &= ~C_INITIALIZED;
5350 /** Push a page onto the top of the cursor's stack.
5351 * Set #MDB_TXN_ERROR on failure.
5354 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5356 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5357 DDBI(mc), (void *) mc));
5359 if (mc->mc_snum >= CURSOR_STACK) {
5360 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5361 return MDB_CURSOR_FULL;
5364 mc->mc_top = mc->mc_snum++;
5365 mc->mc_pg[mc->mc_top] = mp;
5366 mc->mc_ki[mc->mc_top] = 0;
5371 /** Find the address of the page corresponding to a given page number.
5372 * Set #MDB_TXN_ERROR on failure.
5373 * @param[in] mc the cursor accessing the page.
5374 * @param[in] pgno the page number for the page to retrieve.
5375 * @param[out] ret address of a pointer where the page's address will be stored.
5376 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5377 * @return 0 on success, non-zero on failure.
5380 mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **ret, int *lvl)
5382 MDB_txn *txn = mc->mc_txn;
5383 MDB_env *env = txn->mt_env;
5387 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5391 MDB_ID2L dl = tx2->mt_u.dirty_list;
5393 /* Spilled pages were dirtied in this txn and flushed
5394 * because the dirty list got full. Bring this page
5395 * back in from the map (but don't unspill it here,
5396 * leave that unless page_touch happens again).
5398 if (tx2->mt_spill_pgs) {
5399 MDB_ID pn = pgno << 1;
5400 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5401 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5402 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5407 unsigned x = mdb_mid2l_search(dl, pgno);
5408 if (x <= dl[0].mid && dl[x].mid == pgno) {
5414 } while ((tx2 = tx2->mt_parent) != NULL);
5417 if (pgno < txn->mt_next_pgno) {
5419 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5421 DPRINTF(("page %"Z"u not found", pgno));
5422 txn->mt_flags |= MDB_TXN_ERROR;
5423 return MDB_PAGE_NOTFOUND;
5433 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5434 * The cursor is at the root page, set up the rest of it.
5437 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5439 MDB_page *mp = mc->mc_pg[mc->mc_top];
5443 while (IS_BRANCH(mp)) {
5447 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5448 /* Don't assert on branch pages in the FreeDB. We can get here
5449 * while in the process of rebalancing a FreeDB branch page; we must
5450 * let that proceed. ITS#8336
5452 mdb_cassert(mc, !mc->mc_dbi || NUMKEYS(mp) > 1);
5453 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5455 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5457 if (flags & MDB_PS_LAST)
5458 i = NUMKEYS(mp) - 1;
5461 node = mdb_node_search(mc, key, &exact);
5463 i = NUMKEYS(mp) - 1;
5465 i = mc->mc_ki[mc->mc_top];
5467 mdb_cassert(mc, i > 0);
5471 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5474 mdb_cassert(mc, i < NUMKEYS(mp));
5475 node = NODEPTR(mp, i);
5477 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
5480 mc->mc_ki[mc->mc_top] = i;
5481 if ((rc = mdb_cursor_push(mc, mp)))
5484 if (flags & MDB_PS_MODIFY) {
5485 if ((rc = mdb_page_touch(mc)) != 0)
5487 mp = mc->mc_pg[mc->mc_top];
5492 DPRINTF(("internal error, index points to a %02X page!?",
5494 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5495 return MDB_CORRUPTED;
5498 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5499 key ? DKEY(key) : "null"));
5500 mc->mc_flags |= C_INITIALIZED;
5501 mc->mc_flags &= ~C_EOF;
5506 /** Search for the lowest key under the current branch page.
5507 * This just bypasses a NUMKEYS check in the current page
5508 * before calling mdb_page_search_root(), because the callers
5509 * are all in situations where the current page is known to
5513 mdb_page_search_lowest(MDB_cursor *mc)
5515 MDB_page *mp = mc->mc_pg[mc->mc_top];
5516 MDB_node *node = NODEPTR(mp, 0);
5519 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
5522 mc->mc_ki[mc->mc_top] = 0;
5523 if ((rc = mdb_cursor_push(mc, mp)))
5525 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5528 /** Search for the page a given key should be in.
5529 * Push it and its parent pages on the cursor stack.
5530 * @param[in,out] mc the cursor for this operation.
5531 * @param[in] key the key to search for, or NULL for first/last page.
5532 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5533 * are touched (updated with new page numbers).
5534 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5535 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5536 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5537 * @return 0 on success, non-zero on failure.
5540 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5545 /* Make sure the txn is still viable, then find the root from
5546 * the txn's db table and set it as the root of the cursor's stack.
5548 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
5549 DPUTS("transaction may not be used now");
5552 /* Make sure we're using an up-to-date root */
5553 if (*mc->mc_dbflag & DB_STALE) {
5555 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5557 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5558 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5565 MDB_node *leaf = mdb_node_search(&mc2,
5566 &mc->mc_dbx->md_name, &exact);
5568 return MDB_NOTFOUND;
5569 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5570 return MDB_INCOMPATIBLE; /* not a named DB */
5571 rc = mdb_node_read(&mc2, leaf, &data);
5574 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5576 /* The txn may not know this DBI, or another process may
5577 * have dropped and recreated the DB with other flags.
5579 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5580 return MDB_INCOMPATIBLE;
5581 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5583 *mc->mc_dbflag &= ~DB_STALE;
5585 root = mc->mc_db->md_root;
5587 if (root == P_INVALID) { /* Tree is empty. */
5588 DPUTS("tree is empty");
5589 return MDB_NOTFOUND;
5593 mdb_cassert(mc, root > 1);
5594 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5595 if ((rc = mdb_page_get(mc, root, &mc->mc_pg[0], NULL)) != 0)
5601 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5602 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5604 if (flags & MDB_PS_MODIFY) {
5605 if ((rc = mdb_page_touch(mc)))
5609 if (flags & MDB_PS_ROOTONLY)
5612 return mdb_page_search_root(mc, key, flags);
5616 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5618 MDB_txn *txn = mc->mc_txn;
5619 pgno_t pg = mp->mp_pgno;
5620 unsigned x = 0, ovpages = mp->mp_pages;
5621 MDB_env *env = txn->mt_env;
5622 MDB_IDL sl = txn->mt_spill_pgs;
5623 MDB_ID pn = pg << 1;
5626 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5627 /* If the page is dirty or on the spill list we just acquired it,
5628 * so we should give it back to our current free list, if any.
5629 * Otherwise put it onto the list of pages we freed in this txn.
5631 * Won't create me_pghead: me_pglast must be inited along with it.
5632 * Unsupported in nested txns: They would need to hide the page
5633 * range in ancestor txns' dirty and spilled lists.
5635 if (env->me_pghead &&
5637 ((mp->mp_flags & P_DIRTY) ||
5638 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5642 MDB_ID2 *dl, ix, iy;
5643 rc = mdb_midl_need(&env->me_pghead, ovpages);
5646 if (!(mp->mp_flags & P_DIRTY)) {
5647 /* This page is no longer spilled */
5654 /* Remove from dirty list */
5655 dl = txn->mt_u.dirty_list;
5657 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5663 mdb_cassert(mc, x > 1);
5665 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5666 txn->mt_flags |= MDB_TXN_ERROR;
5667 return MDB_CORRUPTED;
5670 txn->mt_dirty_room++;
5671 if (!(env->me_flags & MDB_WRITEMAP))
5672 mdb_dpage_free(env, mp);
5674 /* Insert in me_pghead */
5675 mop = env->me_pghead;
5676 j = mop[0] + ovpages;
5677 for (i = mop[0]; i && mop[i] < pg; i--)
5683 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5687 mc->mc_db->md_overflow_pages -= ovpages;
5691 /** Return the data associated with a given node.
5692 * @param[in] mc The cursor for this operation.
5693 * @param[in] leaf The node being read.
5694 * @param[out] data Updated to point to the node's data.
5695 * @return 0 on success, non-zero on failure.
5698 mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data)
5700 MDB_page *omp; /* overflow page */
5704 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5705 data->mv_size = NODEDSZ(leaf);
5706 data->mv_data = NODEDATA(leaf);
5710 /* Read overflow data.
5712 data->mv_size = NODEDSZ(leaf);
5713 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5714 if ((rc = mdb_page_get(mc, pgno, &omp, NULL)) != 0) {
5715 DPRINTF(("read overflow page %"Z"u failed", pgno));
5718 data->mv_data = METADATA(omp);
5724 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5725 MDB_val *key, MDB_val *data)
5732 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5734 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
5737 if (txn->mt_flags & MDB_TXN_BLOCKED)
5740 mdb_cursor_init(&mc, txn, dbi, &mx);
5741 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5744 /** Find a sibling for a page.
5745 * Replaces the page at the top of the cursor's stack with the
5746 * specified sibling, if one exists.
5747 * @param[in] mc The cursor for this operation.
5748 * @param[in] move_right Non-zero if the right sibling is requested,
5749 * otherwise the left sibling.
5750 * @return 0 on success, non-zero on failure.
5753 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5759 if (mc->mc_snum < 2) {
5760 return MDB_NOTFOUND; /* root has no siblings */
5764 DPRINTF(("parent page is page %"Z"u, index %u",
5765 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5767 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5768 : (mc->mc_ki[mc->mc_top] == 0)) {
5769 DPRINTF(("no more keys left, moving to %s sibling",
5770 move_right ? "right" : "left"));
5771 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5772 /* undo cursor_pop before returning */
5779 mc->mc_ki[mc->mc_top]++;
5781 mc->mc_ki[mc->mc_top]--;
5782 DPRINTF(("just moving to %s index key %u",
5783 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5785 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5787 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5788 if ((rc = mdb_page_get(mc, NODEPGNO(indx), &mp, NULL)) != 0) {
5789 /* mc will be inconsistent if caller does mc_snum++ as above */
5790 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5794 mdb_cursor_push(mc, mp);
5796 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5801 /** Move the cursor to the next data item. */
5803 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5809 if ((mc->mc_flags & C_EOF) ||
5810 ((mc->mc_flags & C_DEL) && op == MDB_NEXT_DUP)) {
5811 return MDB_NOTFOUND;
5813 if (!(mc->mc_flags & C_INITIALIZED))
5814 return mdb_cursor_first(mc, key, data);
5816 mp = mc->mc_pg[mc->mc_top];
5818 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5819 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5820 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5821 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5822 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5823 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5824 if (rc == MDB_SUCCESS)
5825 MDB_GET_KEY(leaf, key);
5830 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5831 if (op == MDB_NEXT_DUP)
5832 return MDB_NOTFOUND;
5836 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5837 mdb_dbg_pgno(mp), (void *) mc));
5838 if (mc->mc_flags & C_DEL) {
5839 mc->mc_flags ^= C_DEL;
5843 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5844 DPUTS("=====> move to next sibling page");
5845 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5846 mc->mc_flags |= C_EOF;
5849 mp = mc->mc_pg[mc->mc_top];
5850 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5852 mc->mc_ki[mc->mc_top]++;
5855 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5856 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5859 key->mv_size = mc->mc_db->md_pad;
5860 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5864 mdb_cassert(mc, IS_LEAF(mp));
5865 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5867 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5868 mdb_xcursor_init1(mc, leaf);
5871 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
5874 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5875 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5876 if (rc != MDB_SUCCESS)
5881 MDB_GET_KEY(leaf, key);
5885 /** Move the cursor to the previous data item. */
5887 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5893 if (!(mc->mc_flags & C_INITIALIZED)) {
5894 rc = mdb_cursor_last(mc, key, data);
5897 mc->mc_ki[mc->mc_top]++;
5900 mp = mc->mc_pg[mc->mc_top];
5902 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5903 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5904 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5905 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5906 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5907 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5908 if (rc == MDB_SUCCESS) {
5909 MDB_GET_KEY(leaf, key);
5910 mc->mc_flags &= ~C_EOF;
5916 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5917 if (op == MDB_PREV_DUP)
5918 return MDB_NOTFOUND;
5922 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5923 mdb_dbg_pgno(mp), (void *) mc));
5925 mc->mc_flags &= ~(C_EOF|C_DEL);
5927 if (mc->mc_ki[mc->mc_top] == 0) {
5928 DPUTS("=====> move to prev sibling page");
5929 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5932 mp = mc->mc_pg[mc->mc_top];
5933 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5934 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5936 mc->mc_ki[mc->mc_top]--;
5938 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5939 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5942 key->mv_size = mc->mc_db->md_pad;
5943 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5947 mdb_cassert(mc, IS_LEAF(mp));
5948 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5950 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5951 mdb_xcursor_init1(mc, leaf);
5954 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
5957 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5958 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5959 if (rc != MDB_SUCCESS)
5964 MDB_GET_KEY(leaf, key);
5968 /** Set the cursor on a specific data item. */
5970 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5971 MDB_cursor_op op, int *exactp)
5975 MDB_node *leaf = NULL;
5978 if (key->mv_size == 0)
5979 return MDB_BAD_VALSIZE;
5982 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5984 /* See if we're already on the right page */
5985 if (mc->mc_flags & C_INITIALIZED) {
5988 mp = mc->mc_pg[mc->mc_top];
5990 mc->mc_ki[mc->mc_top] = 0;
5991 return MDB_NOTFOUND;
5993 if (mp->mp_flags & P_LEAF2) {
5994 nodekey.mv_size = mc->mc_db->md_pad;
5995 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5997 leaf = NODEPTR(mp, 0);
5998 MDB_GET_KEY2(leaf, nodekey);
6000 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6002 /* Probably happens rarely, but first node on the page
6003 * was the one we wanted.
6005 mc->mc_ki[mc->mc_top] = 0;
6012 unsigned int nkeys = NUMKEYS(mp);
6014 if (mp->mp_flags & P_LEAF2) {
6015 nodekey.mv_data = LEAF2KEY(mp,
6016 nkeys-1, nodekey.mv_size);
6018 leaf = NODEPTR(mp, nkeys-1);
6019 MDB_GET_KEY2(leaf, nodekey);
6021 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6023 /* last node was the one we wanted */
6024 mc->mc_ki[mc->mc_top] = nkeys-1;
6030 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
6031 /* This is definitely the right page, skip search_page */
6032 if (mp->mp_flags & P_LEAF2) {
6033 nodekey.mv_data = LEAF2KEY(mp,
6034 mc->mc_ki[mc->mc_top], nodekey.mv_size);
6036 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6037 MDB_GET_KEY2(leaf, nodekey);
6039 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6041 /* current node was the one we wanted */
6048 mc->mc_flags &= ~C_EOF;
6052 /* If any parents have right-sibs, search.
6053 * Otherwise, there's nothing further.
6055 for (i=0; i<mc->mc_top; i++)
6057 NUMKEYS(mc->mc_pg[i])-1)
6059 if (i == mc->mc_top) {
6060 /* There are no other pages */
6061 mc->mc_ki[mc->mc_top] = nkeys;
6062 return MDB_NOTFOUND;
6066 /* There are no other pages */
6067 mc->mc_ki[mc->mc_top] = 0;
6068 if (op == MDB_SET_RANGE && !exactp) {
6072 return MDB_NOTFOUND;
6078 rc = mdb_page_search(mc, key, 0);
6079 if (rc != MDB_SUCCESS)
6082 mp = mc->mc_pg[mc->mc_top];
6083 mdb_cassert(mc, IS_LEAF(mp));
6086 leaf = mdb_node_search(mc, key, exactp);
6087 if (exactp != NULL && !*exactp) {
6088 /* MDB_SET specified and not an exact match. */
6089 return MDB_NOTFOUND;
6093 DPUTS("===> inexact leaf not found, goto sibling");
6094 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
6095 mc->mc_flags |= C_EOF;
6096 return rc; /* no entries matched */
6098 mp = mc->mc_pg[mc->mc_top];
6099 mdb_cassert(mc, IS_LEAF(mp));
6100 leaf = NODEPTR(mp, 0);
6104 mc->mc_flags |= C_INITIALIZED;
6105 mc->mc_flags &= ~C_EOF;
6108 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
6109 key->mv_size = mc->mc_db->md_pad;
6110 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6115 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6116 mdb_xcursor_init1(mc, leaf);
6119 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6120 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
6121 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6124 if (op == MDB_GET_BOTH) {
6130 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
6131 if (rc != MDB_SUCCESS)
6134 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
6137 if ((rc = mdb_node_read(mc, leaf, &olddata)) != MDB_SUCCESS)
6139 dcmp = mc->mc_dbx->md_dcmp;
6140 #if UINT_MAX < SIZE_MAX
6141 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6142 dcmp = mdb_cmp_clong;
6144 rc = dcmp(data, &olddata);
6146 if (op == MDB_GET_BOTH || rc > 0)
6147 return MDB_NOTFOUND;
6154 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6155 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6160 /* The key already matches in all other cases */
6161 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
6162 MDB_GET_KEY(leaf, key);
6163 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
6168 /** Move the cursor to the first item in the database. */
6170 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6176 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6178 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6179 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
6180 if (rc != MDB_SUCCESS)
6183 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6185 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
6186 mc->mc_flags |= C_INITIALIZED;
6187 mc->mc_flags &= ~C_EOF;
6189 mc->mc_ki[mc->mc_top] = 0;
6191 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6192 key->mv_size = mc->mc_db->md_pad;
6193 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
6198 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6199 mdb_xcursor_init1(mc, leaf);
6200 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6204 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6208 MDB_GET_KEY(leaf, key);
6212 /** Move the cursor to the last item in the database. */
6214 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6220 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6222 if (!(mc->mc_flags & C_EOF)) {
6224 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6225 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
6226 if (rc != MDB_SUCCESS)
6229 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6232 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6233 mc->mc_flags |= C_INITIALIZED|C_EOF;
6234 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6236 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6237 key->mv_size = mc->mc_db->md_pad;
6238 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6243 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6244 mdb_xcursor_init1(mc, leaf);
6245 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6249 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6254 MDB_GET_KEY(leaf, key);
6259 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6264 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6269 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
6273 case MDB_GET_CURRENT:
6274 if (!(mc->mc_flags & C_INITIALIZED)) {
6277 MDB_page *mp = mc->mc_pg[mc->mc_top];
6278 int nkeys = NUMKEYS(mp);
6279 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6280 mc->mc_ki[mc->mc_top] = nkeys;
6286 key->mv_size = mc->mc_db->md_pad;
6287 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6289 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6290 MDB_GET_KEY(leaf, key);
6292 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6293 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6295 rc = mdb_node_read(mc, leaf, data);
6302 case MDB_GET_BOTH_RANGE:
6307 if (mc->mc_xcursor == NULL) {
6308 rc = MDB_INCOMPATIBLE;
6318 rc = mdb_cursor_set(mc, key, data, op,
6319 op == MDB_SET_RANGE ? NULL : &exact);
6322 case MDB_GET_MULTIPLE:
6323 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6327 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6328 rc = MDB_INCOMPATIBLE;
6332 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6333 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6336 case MDB_NEXT_MULTIPLE:
6341 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6342 rc = MDB_INCOMPATIBLE;
6345 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6346 if (rc == MDB_SUCCESS) {
6347 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6350 mx = &mc->mc_xcursor->mx_cursor;
6351 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6353 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6354 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6360 case MDB_PREV_MULTIPLE:
6365 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6366 rc = MDB_INCOMPATIBLE;
6369 if (!(mc->mc_flags & C_INITIALIZED))
6370 rc = mdb_cursor_last(mc, key, data);
6373 if (rc == MDB_SUCCESS) {
6374 MDB_cursor *mx = &mc->mc_xcursor->mx_cursor;
6375 if (mx->mc_flags & C_INITIALIZED) {
6376 rc = mdb_cursor_sibling(mx, 0);
6377 if (rc == MDB_SUCCESS)
6386 case MDB_NEXT_NODUP:
6387 rc = mdb_cursor_next(mc, key, data, op);
6391 case MDB_PREV_NODUP:
6392 rc = mdb_cursor_prev(mc, key, data, op);
6395 rc = mdb_cursor_first(mc, key, data);
6398 mfunc = mdb_cursor_first;
6400 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6404 if (mc->mc_xcursor == NULL) {
6405 rc = MDB_INCOMPATIBLE;
6409 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6410 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6411 MDB_GET_KEY(leaf, key);
6412 rc = mdb_node_read(mc, leaf, data);
6416 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6420 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6423 rc = mdb_cursor_last(mc, key, data);
6426 mfunc = mdb_cursor_last;
6429 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6434 if (mc->mc_flags & C_DEL)
6435 mc->mc_flags ^= C_DEL;
6440 /** Touch all the pages in the cursor stack. Set mc_top.
6441 * Makes sure all the pages are writable, before attempting a write operation.
6442 * @param[in] mc The cursor to operate on.
6445 mdb_cursor_touch(MDB_cursor *mc)
6447 int rc = MDB_SUCCESS;
6449 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & (DB_DIRTY|DB_DUPDATA))) {
6450 /* Touch DB record of named DB */
6453 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6455 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6456 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6459 *mc->mc_dbflag |= DB_DIRTY;
6464 rc = mdb_page_touch(mc);
6465 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6466 mc->mc_top = mc->mc_snum-1;
6471 /** Do not spill pages to disk if txn is getting full, may fail instead */
6472 #define MDB_NOSPILL 0x8000
6475 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6479 MDB_node *leaf = NULL;
6480 MDB_page *fp, *mp, *sub_root = NULL;
6482 MDB_val xdata, *rdata, dkey, olddata;
6484 int do_sub = 0, insert_key, insert_data;
6485 unsigned int mcount = 0, dcount = 0, nospill;
6488 unsigned int nflags;
6491 if (mc == NULL || key == NULL)
6494 env = mc->mc_txn->mt_env;
6496 /* Check this first so counter will always be zero on any
6499 if (flags & MDB_MULTIPLE) {
6500 dcount = data[1].mv_size;
6501 data[1].mv_size = 0;
6502 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6503 return MDB_INCOMPATIBLE;
6506 nospill = flags & MDB_NOSPILL;
6507 flags &= ~MDB_NOSPILL;
6509 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6510 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6512 if (key->mv_size-1 >= ENV_MAXKEY(env))
6513 return MDB_BAD_VALSIZE;
6515 #if SIZE_MAX > MAXDATASIZE
6516 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6517 return MDB_BAD_VALSIZE;
6519 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6520 return MDB_BAD_VALSIZE;
6523 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6524 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6528 if (flags == MDB_CURRENT) {
6529 if (!(mc->mc_flags & C_INITIALIZED))
6532 } else if (mc->mc_db->md_root == P_INVALID) {
6533 /* new database, cursor has nothing to point to */
6536 mc->mc_flags &= ~C_INITIALIZED;
6541 if (flags & MDB_APPEND) {
6543 rc = mdb_cursor_last(mc, &k2, &d2);
6545 rc = mc->mc_dbx->md_cmp(key, &k2);
6548 mc->mc_ki[mc->mc_top]++;
6550 /* new key is <= last key */
6555 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6557 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6558 DPRINTF(("duplicate key [%s]", DKEY(key)));
6560 return MDB_KEYEXIST;
6562 if (rc && rc != MDB_NOTFOUND)
6566 if (mc->mc_flags & C_DEL)
6567 mc->mc_flags ^= C_DEL;
6569 /* Cursor is positioned, check for room in the dirty list */
6571 if (flags & MDB_MULTIPLE) {
6573 xdata.mv_size = data->mv_size * dcount;
6577 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6581 if (rc == MDB_NO_ROOT) {
6583 /* new database, write a root leaf page */
6584 DPUTS("allocating new root leaf page");
6585 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6588 mdb_cursor_push(mc, np);
6589 mc->mc_db->md_root = np->mp_pgno;
6590 mc->mc_db->md_depth++;
6591 *mc->mc_dbflag |= DB_DIRTY;
6592 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6594 np->mp_flags |= P_LEAF2;
6595 mc->mc_flags |= C_INITIALIZED;
6597 /* make sure all cursor pages are writable */
6598 rc2 = mdb_cursor_touch(mc);
6603 insert_key = insert_data = rc;
6605 /* The key does not exist */
6606 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6607 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6608 LEAFSIZE(key, data) > env->me_nodemax)
6610 /* Too big for a node, insert in sub-DB. Set up an empty
6611 * "old sub-page" for prep_subDB to expand to a full page.
6613 fp_flags = P_LEAF|P_DIRTY;
6615 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6616 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6617 olddata.mv_size = PAGEHDRSZ;
6621 /* there's only a key anyway, so this is a no-op */
6622 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6624 unsigned int ksize = mc->mc_db->md_pad;
6625 if (key->mv_size != ksize)
6626 return MDB_BAD_VALSIZE;
6627 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6628 memcpy(ptr, key->mv_data, ksize);
6630 /* if overwriting slot 0 of leaf, need to
6631 * update branch key if there is a parent page
6633 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6634 unsigned short dtop = 1;
6636 /* slot 0 is always an empty key, find real slot */
6637 while (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6641 if (mc->mc_ki[mc->mc_top])
6642 rc2 = mdb_update_key(mc, key);
6653 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6654 olddata.mv_size = NODEDSZ(leaf);
6655 olddata.mv_data = NODEDATA(leaf);
6658 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6659 /* Prepare (sub-)page/sub-DB to accept the new item,
6660 * if needed. fp: old sub-page or a header faking
6661 * it. mp: new (sub-)page. offset: growth in page
6662 * size. xdata: node data with new page or DB.
6664 unsigned i, offset = 0;
6665 mp = fp = xdata.mv_data = env->me_pbuf;
6666 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6668 /* Was a single item before, must convert now */
6669 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6671 /* Just overwrite the current item */
6672 if (flags == MDB_CURRENT)
6674 dcmp = mc->mc_dbx->md_dcmp;
6675 #if UINT_MAX < SIZE_MAX
6676 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6677 dcmp = mdb_cmp_clong;
6679 /* does data match? */
6680 if (!dcmp(data, &olddata)) {
6681 if (flags & (MDB_NODUPDATA|MDB_APPENDDUP))
6682 return MDB_KEYEXIST;
6687 /* Back up original data item */
6688 dkey.mv_size = olddata.mv_size;
6689 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6691 /* Make sub-page header for the dup items, with dummy body */
6692 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6693 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6694 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6695 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6696 fp->mp_flags |= P_LEAF2;
6697 fp->mp_pad = data->mv_size;
6698 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6700 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6701 (dkey.mv_size & 1) + (data->mv_size & 1);
6703 fp->mp_upper = xdata.mv_size - PAGEBASE;
6704 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6705 } else if (leaf->mn_flags & F_SUBDATA) {
6706 /* Data is on sub-DB, just store it */
6707 flags |= F_DUPDATA|F_SUBDATA;
6710 /* Data is on sub-page */
6711 fp = olddata.mv_data;
6714 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6715 offset = EVEN(NODESIZE + sizeof(indx_t) +
6719 offset = fp->mp_pad;
6720 if (SIZELEFT(fp) < offset) {
6721 offset *= 4; /* space for 4 more */
6724 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6726 fp->mp_flags |= P_DIRTY;
6727 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6728 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6732 xdata.mv_size = olddata.mv_size + offset;
6735 fp_flags = fp->mp_flags;
6736 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6737 /* Too big for a sub-page, convert to sub-DB */
6738 fp_flags &= ~P_SUBP;
6740 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6741 fp_flags |= P_LEAF2;
6742 dummy.md_pad = fp->mp_pad;
6743 dummy.md_flags = MDB_DUPFIXED;
6744 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6745 dummy.md_flags |= MDB_INTEGERKEY;
6751 dummy.md_branch_pages = 0;
6752 dummy.md_leaf_pages = 1;
6753 dummy.md_overflow_pages = 0;
6754 dummy.md_entries = NUMKEYS(fp);
6755 xdata.mv_size = sizeof(MDB_db);
6756 xdata.mv_data = &dummy;
6757 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6759 offset = env->me_psize - olddata.mv_size;
6760 flags |= F_DUPDATA|F_SUBDATA;
6761 dummy.md_root = mp->mp_pgno;
6765 mp->mp_flags = fp_flags | P_DIRTY;
6766 mp->mp_pad = fp->mp_pad;
6767 mp->mp_lower = fp->mp_lower;
6768 mp->mp_upper = fp->mp_upper + offset;
6769 if (fp_flags & P_LEAF2) {
6770 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6772 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6773 olddata.mv_size - fp->mp_upper - PAGEBASE);
6774 for (i=0; i<NUMKEYS(fp); i++)
6775 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6783 mdb_node_del(mc, 0);
6787 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6788 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6789 return MDB_INCOMPATIBLE;
6790 /* overflow page overwrites need special handling */
6791 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6794 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6796 memcpy(&pg, olddata.mv_data, sizeof(pg));
6797 if ((rc2 = mdb_page_get(mc, pg, &omp, &level)) != 0)
6799 ovpages = omp->mp_pages;
6801 /* Is the ov page large enough? */
6802 if (ovpages >= dpages) {
6803 if (!(omp->mp_flags & P_DIRTY) &&
6804 (level || (env->me_flags & MDB_WRITEMAP)))
6806 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6809 level = 0; /* dirty in this txn or clean */
6812 if (omp->mp_flags & P_DIRTY) {
6813 /* yes, overwrite it. Note in this case we don't
6814 * bother to try shrinking the page if the new data
6815 * is smaller than the overflow threshold.
6818 /* It is writable only in a parent txn */
6819 size_t sz = (size_t) env->me_psize * ovpages, off;
6820 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6826 /* Note - this page is already counted in parent's dirty_room */
6827 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6828 mdb_cassert(mc, rc2 == 0);
6829 /* Currently we make the page look as with put() in the
6830 * parent txn, in case the user peeks at MDB_RESERVEd
6831 * or unused parts. Some users treat ovpages specially.
6833 if (!(flags & MDB_RESERVE)) {
6834 /* Skip the part where LMDB will put *data.
6835 * Copy end of page, adjusting alignment so
6836 * compiler may copy words instead of bytes.
6838 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6839 memcpy((size_t *)((char *)np + off),
6840 (size_t *)((char *)omp + off), sz - off);
6843 memcpy(np, omp, sz); /* Copy beginning of page */
6846 SETDSZ(leaf, data->mv_size);
6847 if (F_ISSET(flags, MDB_RESERVE))
6848 data->mv_data = METADATA(omp);
6850 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6854 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6856 } else if (data->mv_size == olddata.mv_size) {
6857 /* same size, just replace it. Note that we could
6858 * also reuse this node if the new data is smaller,
6859 * but instead we opt to shrink the node in that case.
6861 if (F_ISSET(flags, MDB_RESERVE))
6862 data->mv_data = olddata.mv_data;
6863 else if (!(mc->mc_flags & C_SUB))
6864 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6866 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6871 mdb_node_del(mc, 0);
6877 nflags = flags & NODE_ADD_FLAGS;
6878 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6879 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6880 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6881 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6883 nflags |= MDB_SPLIT_REPLACE;
6884 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6886 /* There is room already in this leaf page. */
6887 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6889 /* Adjust other cursors pointing to mp */
6890 MDB_cursor *m2, *m3;
6891 MDB_dbi dbi = mc->mc_dbi;
6892 unsigned i = mc->mc_top;
6893 MDB_page *mp = mc->mc_pg[i];
6895 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6896 if (mc->mc_flags & C_SUB)
6897 m3 = &m2->mc_xcursor->mx_cursor;
6900 if (m3 == mc || m3->mc_snum < mc->mc_snum || m3->mc_pg[i] != mp) continue;
6901 if (m3->mc_ki[i] >= mc->mc_ki[i] && insert_key) {
6904 if (XCURSOR_INITED(m3))
6905 XCURSOR_REFRESH(m3, mp, m3->mc_ki[i]);
6910 if (rc == MDB_SUCCESS) {
6911 /* Now store the actual data in the child DB. Note that we're
6912 * storing the user data in the keys field, so there are strict
6913 * size limits on dupdata. The actual data fields of the child
6914 * DB are all zero size.
6917 int xflags, new_dupdata;
6922 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6923 if (flags & MDB_CURRENT) {
6924 xflags = MDB_CURRENT|MDB_NOSPILL;
6926 mdb_xcursor_init1(mc, leaf);
6927 xflags = (flags & MDB_NODUPDATA) ?
6928 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6931 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
6932 new_dupdata = (int)dkey.mv_size;
6933 /* converted, write the original data first */
6935 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6938 /* we've done our job */
6941 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
6942 /* Adjust other cursors pointing to mp */
6944 MDB_xcursor *mx = mc->mc_xcursor;
6945 unsigned i = mc->mc_top;
6946 MDB_page *mp = mc->mc_pg[i];
6947 int nkeys = NUMKEYS(mp);
6949 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6950 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6951 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6952 if (m2->mc_pg[i] == mp) {
6953 if (m2->mc_ki[i] == mc->mc_ki[i]) {
6954 mdb_xcursor_init2(m2, mx, new_dupdata);
6955 } else if (!insert_key && m2->mc_ki[i] < nkeys) {
6956 XCURSOR_REFRESH(m2, mp, m2->mc_ki[i]);
6961 ecount = mc->mc_xcursor->mx_db.md_entries;
6962 if (flags & MDB_APPENDDUP)
6963 xflags |= MDB_APPEND;
6964 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6965 if (flags & F_SUBDATA) {
6966 void *db = NODEDATA(leaf);
6967 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6969 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6971 /* Increment count unless we just replaced an existing item. */
6973 mc->mc_db->md_entries++;
6975 /* Invalidate txn if we created an empty sub-DB */
6978 /* If we succeeded and the key didn't exist before,
6979 * make sure the cursor is marked valid.
6981 mc->mc_flags |= C_INITIALIZED;
6983 if (flags & MDB_MULTIPLE) {
6986 /* let caller know how many succeeded, if any */
6987 data[1].mv_size = mcount;
6988 if (mcount < dcount) {
6989 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6990 insert_key = insert_data = 0;
6997 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
7000 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7005 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
7011 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
7012 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7014 if (!(mc->mc_flags & C_INITIALIZED))
7017 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
7018 return MDB_NOTFOUND;
7020 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
7023 rc = mdb_cursor_touch(mc);
7027 mp = mc->mc_pg[mc->mc_top];
7030 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7032 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7033 if (flags & MDB_NODUPDATA) {
7034 /* mdb_cursor_del0() will subtract the final entry */
7035 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
7036 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7038 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
7039 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7041 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
7044 /* If sub-DB still has entries, we're done */
7045 if (mc->mc_xcursor->mx_db.md_entries) {
7046 if (leaf->mn_flags & F_SUBDATA) {
7047 /* update subDB info */
7048 void *db = NODEDATA(leaf);
7049 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
7052 /* shrink fake page */
7053 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
7054 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7055 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7056 /* fix other sub-DB cursors pointed at fake pages on this page */
7057 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
7058 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
7059 if (!(m2->mc_flags & C_INITIALIZED)) continue;
7060 if (m2->mc_pg[mc->mc_top] == mp) {
7061 MDB_node *n2 = leaf;
7062 if (m2->mc_ki[mc->mc_top] != mc->mc_ki[mc->mc_top]) {
7063 n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
7064 if (n2->mn_flags & F_SUBDATA) continue;
7066 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
7070 mc->mc_db->md_entries--;
7073 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7075 /* otherwise fall thru and delete the sub-DB */
7078 if (leaf->mn_flags & F_SUBDATA) {
7079 /* add all the child DB's pages to the free list */
7080 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7085 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
7086 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
7087 rc = MDB_INCOMPATIBLE;
7091 /* add overflow pages to free list */
7092 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7096 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7097 if ((rc = mdb_page_get(mc, pg, &omp, NULL)) ||
7098 (rc = mdb_ovpage_free(mc, omp)))
7103 return mdb_cursor_del0(mc);
7106 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7110 /** Allocate and initialize new pages for a database.
7111 * Set #MDB_TXN_ERROR on failure.
7112 * @param[in] mc a cursor on the database being added to.
7113 * @param[in] flags flags defining what type of page is being allocated.
7114 * @param[in] num the number of pages to allocate. This is usually 1,
7115 * unless allocating overflow pages for a large record.
7116 * @param[out] mp Address of a page, or NULL on failure.
7117 * @return 0 on success, non-zero on failure.
7120 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
7125 if ((rc = mdb_page_alloc(mc, num, &np)))
7127 DPRINTF(("allocated new mpage %"Z"u, page size %u",
7128 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
7129 np->mp_flags = flags | P_DIRTY;
7130 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
7131 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
7134 mc->mc_db->md_branch_pages++;
7135 else if (IS_LEAF(np))
7136 mc->mc_db->md_leaf_pages++;
7137 else if (IS_OVERFLOW(np)) {
7138 mc->mc_db->md_overflow_pages += num;
7146 /** Calculate the size of a leaf node.
7147 * The size depends on the environment's page size; if a data item
7148 * is too large it will be put onto an overflow page and the node
7149 * size will only include the key and not the data. Sizes are always
7150 * rounded up to an even number of bytes, to guarantee 2-byte alignment
7151 * of the #MDB_node headers.
7152 * @param[in] env The environment handle.
7153 * @param[in] key The key for the node.
7154 * @param[in] data The data for the node.
7155 * @return The number of bytes needed to store the node.
7158 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
7162 sz = LEAFSIZE(key, data);
7163 if (sz > env->me_nodemax) {
7164 /* put on overflow page */
7165 sz -= data->mv_size - sizeof(pgno_t);
7168 return EVEN(sz + sizeof(indx_t));
7171 /** Calculate the size of a branch node.
7172 * The size should depend on the environment's page size but since
7173 * we currently don't support spilling large keys onto overflow
7174 * pages, it's simply the size of the #MDB_node header plus the
7175 * size of the key. Sizes are always rounded up to an even number
7176 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
7177 * @param[in] env The environment handle.
7178 * @param[in] key The key for the node.
7179 * @return The number of bytes needed to store the node.
7182 mdb_branch_size(MDB_env *env, MDB_val *key)
7187 if (sz > env->me_nodemax) {
7188 /* put on overflow page */
7189 /* not implemented */
7190 /* sz -= key->size - sizeof(pgno_t); */
7193 return sz + sizeof(indx_t);
7196 /** Add a node to the page pointed to by the cursor.
7197 * Set #MDB_TXN_ERROR on failure.
7198 * @param[in] mc The cursor for this operation.
7199 * @param[in] indx The index on the page where the new node should be added.
7200 * @param[in] key The key for the new node.
7201 * @param[in] data The data for the new node, if any.
7202 * @param[in] pgno The page number, if adding a branch node.
7203 * @param[in] flags Flags for the node.
7204 * @return 0 on success, non-zero on failure. Possible errors are:
7206 * <li>ENOMEM - failed to allocate overflow pages for the node.
7207 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
7208 * should never happen since all callers already calculate the
7209 * page's free space before calling this function.
7213 mdb_node_add(MDB_cursor *mc, indx_t indx,
7214 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
7217 size_t node_size = NODESIZE;
7221 MDB_page *mp = mc->mc_pg[mc->mc_top];
7222 MDB_page *ofp = NULL; /* overflow page */
7226 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
7228 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
7229 IS_LEAF(mp) ? "leaf" : "branch",
7230 IS_SUBP(mp) ? "sub-" : "",
7231 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
7232 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
7235 /* Move higher keys up one slot. */
7236 int ksize = mc->mc_db->md_pad, dif;
7237 char *ptr = LEAF2KEY(mp, indx, ksize);
7238 dif = NUMKEYS(mp) - indx;
7240 memmove(ptr+ksize, ptr, dif*ksize);
7241 /* insert new key */
7242 memcpy(ptr, key->mv_data, ksize);
7244 /* Just using these for counting */
7245 mp->mp_lower += sizeof(indx_t);
7246 mp->mp_upper -= ksize - sizeof(indx_t);
7250 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
7252 node_size += key->mv_size;
7254 mdb_cassert(mc, key && data);
7255 if (F_ISSET(flags, F_BIGDATA)) {
7256 /* Data already on overflow page. */
7257 node_size += sizeof(pgno_t);
7258 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7259 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7261 /* Put data on overflow page. */
7262 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7263 data->mv_size, node_size+data->mv_size));
7264 node_size = EVEN(node_size + sizeof(pgno_t));
7265 if ((ssize_t)node_size > room)
7267 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7269 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
7273 node_size += data->mv_size;
7276 node_size = EVEN(node_size);
7277 if ((ssize_t)node_size > room)
7281 /* Move higher pointers up one slot. */
7282 for (i = NUMKEYS(mp); i > indx; i--)
7283 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7285 /* Adjust free space offsets. */
7286 ofs = mp->mp_upper - node_size;
7287 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7288 mp->mp_ptrs[indx] = ofs;
7290 mp->mp_lower += sizeof(indx_t);
7292 /* Write the node data. */
7293 node = NODEPTR(mp, indx);
7294 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7295 node->mn_flags = flags;
7297 SETDSZ(node,data->mv_size);
7302 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7305 ndata = NODEDATA(node);
7307 if (F_ISSET(flags, F_BIGDATA))
7308 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7309 else if (F_ISSET(flags, MDB_RESERVE))
7310 data->mv_data = ndata;
7312 memcpy(ndata, data->mv_data, data->mv_size);
7314 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7315 ndata = METADATA(ofp);
7316 if (F_ISSET(flags, MDB_RESERVE))
7317 data->mv_data = ndata;
7319 memcpy(ndata, data->mv_data, data->mv_size);
7326 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7327 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7328 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7329 DPRINTF(("node size = %"Z"u", node_size));
7330 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7331 return MDB_PAGE_FULL;
7334 /** Delete the specified node from a page.
7335 * @param[in] mc Cursor pointing to the node to delete.
7336 * @param[in] ksize The size of a node. Only used if the page is
7337 * part of a #MDB_DUPFIXED database.
7340 mdb_node_del(MDB_cursor *mc, int ksize)
7342 MDB_page *mp = mc->mc_pg[mc->mc_top];
7343 indx_t indx = mc->mc_ki[mc->mc_top];
7345 indx_t i, j, numkeys, ptr;
7349 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7350 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7351 numkeys = NUMKEYS(mp);
7352 mdb_cassert(mc, indx < numkeys);
7355 int x = numkeys - 1 - indx;
7356 base = LEAF2KEY(mp, indx, ksize);
7358 memmove(base, base + ksize, x * ksize);
7359 mp->mp_lower -= sizeof(indx_t);
7360 mp->mp_upper += ksize - sizeof(indx_t);
7364 node = NODEPTR(mp, indx);
7365 sz = NODESIZE + node->mn_ksize;
7367 if (F_ISSET(node->mn_flags, F_BIGDATA))
7368 sz += sizeof(pgno_t);
7370 sz += NODEDSZ(node);
7374 ptr = mp->mp_ptrs[indx];
7375 for (i = j = 0; i < numkeys; i++) {
7377 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7378 if (mp->mp_ptrs[i] < ptr)
7379 mp->mp_ptrs[j] += sz;
7384 base = (char *)mp + mp->mp_upper + PAGEBASE;
7385 memmove(base + sz, base, ptr - mp->mp_upper);
7387 mp->mp_lower -= sizeof(indx_t);
7391 /** Compact the main page after deleting a node on a subpage.
7392 * @param[in] mp The main page to operate on.
7393 * @param[in] indx The index of the subpage on the main page.
7396 mdb_node_shrink(MDB_page *mp, indx_t indx)
7401 indx_t delta, nsize, len, ptr;
7404 node = NODEPTR(mp, indx);
7405 sp = (MDB_page *)NODEDATA(node);
7406 delta = SIZELEFT(sp);
7407 nsize = NODEDSZ(node) - delta;
7409 /* Prepare to shift upward, set len = length(subpage part to shift) */
7413 return; /* do not make the node uneven-sized */
7415 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
7416 for (i = NUMKEYS(sp); --i >= 0; )
7417 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7420 sp->mp_upper = sp->mp_lower;
7421 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
7422 SETDSZ(node, nsize);
7424 /* Shift <lower nodes...initial part of subpage> upward */
7425 base = (char *)mp + mp->mp_upper + PAGEBASE;
7426 memmove(base + delta, base, (char *)sp + len - base);
7428 ptr = mp->mp_ptrs[indx];
7429 for (i = NUMKEYS(mp); --i >= 0; ) {
7430 if (mp->mp_ptrs[i] <= ptr)
7431 mp->mp_ptrs[i] += delta;
7433 mp->mp_upper += delta;
7436 /** Initial setup of a sorted-dups cursor.
7437 * Sorted duplicates are implemented as a sub-database for the given key.
7438 * The duplicate data items are actually keys of the sub-database.
7439 * Operations on the duplicate data items are performed using a sub-cursor
7440 * initialized when the sub-database is first accessed. This function does
7441 * the preliminary setup of the sub-cursor, filling in the fields that
7442 * depend only on the parent DB.
7443 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7446 mdb_xcursor_init0(MDB_cursor *mc)
7448 MDB_xcursor *mx = mc->mc_xcursor;
7450 mx->mx_cursor.mc_xcursor = NULL;
7451 mx->mx_cursor.mc_txn = mc->mc_txn;
7452 mx->mx_cursor.mc_db = &mx->mx_db;
7453 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7454 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7455 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7456 mx->mx_cursor.mc_snum = 0;
7457 mx->mx_cursor.mc_top = 0;
7458 mx->mx_cursor.mc_flags = C_SUB;
7459 mx->mx_dbx.md_name.mv_size = 0;
7460 mx->mx_dbx.md_name.mv_data = NULL;
7461 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7462 mx->mx_dbx.md_dcmp = NULL;
7463 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7466 /** Final setup of a sorted-dups cursor.
7467 * Sets up the fields that depend on the data from the main cursor.
7468 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7469 * @param[in] node The data containing the #MDB_db record for the
7470 * sorted-dup database.
7473 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7475 MDB_xcursor *mx = mc->mc_xcursor;
7477 if (node->mn_flags & F_SUBDATA) {
7478 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7479 mx->mx_cursor.mc_pg[0] = 0;
7480 mx->mx_cursor.mc_snum = 0;
7481 mx->mx_cursor.mc_top = 0;
7482 mx->mx_cursor.mc_flags = C_SUB;
7484 MDB_page *fp = NODEDATA(node);
7485 mx->mx_db.md_pad = 0;
7486 mx->mx_db.md_flags = 0;
7487 mx->mx_db.md_depth = 1;
7488 mx->mx_db.md_branch_pages = 0;
7489 mx->mx_db.md_leaf_pages = 1;
7490 mx->mx_db.md_overflow_pages = 0;
7491 mx->mx_db.md_entries = NUMKEYS(fp);
7492 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7493 mx->mx_cursor.mc_snum = 1;
7494 mx->mx_cursor.mc_top = 0;
7495 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7496 mx->mx_cursor.mc_pg[0] = fp;
7497 mx->mx_cursor.mc_ki[0] = 0;
7498 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7499 mx->mx_db.md_flags = MDB_DUPFIXED;
7500 mx->mx_db.md_pad = fp->mp_pad;
7501 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7502 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7505 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7506 mx->mx_db.md_root));
7507 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DUPDATA;
7508 #if UINT_MAX < SIZE_MAX
7509 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7510 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7515 /** Fixup a sorted-dups cursor due to underlying update.
7516 * Sets up some fields that depend on the data from the main cursor.
7517 * Almost the same as init1, but skips initialization steps if the
7518 * xcursor had already been used.
7519 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
7520 * @param[in] src_mx The xcursor of an up-to-date cursor.
7521 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
7524 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
7526 MDB_xcursor *mx = mc->mc_xcursor;
7529 mx->mx_cursor.mc_snum = 1;
7530 mx->mx_cursor.mc_top = 0;
7531 mx->mx_cursor.mc_flags |= C_INITIALIZED;
7532 mx->mx_cursor.mc_ki[0] = 0;
7533 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DUPDATA;
7534 #if UINT_MAX < SIZE_MAX
7535 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
7537 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
7540 mx->mx_db = src_mx->mx_db;
7541 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
7542 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7543 mx->mx_db.md_root));
7546 /** Initialize a cursor for a given transaction and database. */
7548 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7551 mc->mc_backup = NULL;
7554 mc->mc_db = &txn->mt_dbs[dbi];
7555 mc->mc_dbx = &txn->mt_dbxs[dbi];
7556 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7562 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7563 mdb_tassert(txn, mx != NULL);
7564 mc->mc_xcursor = mx;
7565 mdb_xcursor_init0(mc);
7567 mc->mc_xcursor = NULL;
7569 if (*mc->mc_dbflag & DB_STALE) {
7570 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7575 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7578 size_t size = sizeof(MDB_cursor);
7580 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
7583 if (txn->mt_flags & MDB_TXN_BLOCKED)
7586 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7589 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7590 size += sizeof(MDB_xcursor);
7592 if ((mc = malloc(size)) != NULL) {
7593 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7594 if (txn->mt_cursors) {
7595 mc->mc_next = txn->mt_cursors[dbi];
7596 txn->mt_cursors[dbi] = mc;
7597 mc->mc_flags |= C_UNTRACK;
7609 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7611 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
7614 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7617 if (txn->mt_flags & MDB_TXN_BLOCKED)
7620 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7624 /* Return the count of duplicate data items for the current key */
7626 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7630 if (mc == NULL || countp == NULL)
7633 if (mc->mc_xcursor == NULL)
7634 return MDB_INCOMPATIBLE;
7636 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
7639 if (!(mc->mc_flags & C_INITIALIZED))
7642 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7643 return MDB_NOTFOUND;
7645 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7646 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7649 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7652 *countp = mc->mc_xcursor->mx_db.md_entries;
7658 mdb_cursor_close(MDB_cursor *mc)
7660 if (mc && !mc->mc_backup) {
7661 /* remove from txn, if tracked */
7662 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7663 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7664 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7666 *prev = mc->mc_next;
7673 mdb_cursor_txn(MDB_cursor *mc)
7675 if (!mc) return NULL;
7680 mdb_cursor_dbi(MDB_cursor *mc)
7685 /** Replace the key for a branch node with a new key.
7686 * Set #MDB_TXN_ERROR on failure.
7687 * @param[in] mc Cursor pointing to the node to operate on.
7688 * @param[in] key The new key to use.
7689 * @return 0 on success, non-zero on failure.
7692 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7698 int delta, ksize, oksize;
7699 indx_t ptr, i, numkeys, indx;
7702 indx = mc->mc_ki[mc->mc_top];
7703 mp = mc->mc_pg[mc->mc_top];
7704 node = NODEPTR(mp, indx);
7705 ptr = mp->mp_ptrs[indx];
7709 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7710 k2.mv_data = NODEKEY(node);
7711 k2.mv_size = node->mn_ksize;
7712 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7714 mdb_dkey(&k2, kbuf2),
7720 /* Sizes must be 2-byte aligned. */
7721 ksize = EVEN(key->mv_size);
7722 oksize = EVEN(node->mn_ksize);
7723 delta = ksize - oksize;
7725 /* Shift node contents if EVEN(key length) changed. */
7727 if (delta > 0 && SIZELEFT(mp) < delta) {
7729 /* not enough space left, do a delete and split */
7730 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7731 pgno = NODEPGNO(node);
7732 mdb_node_del(mc, 0);
7733 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7736 numkeys = NUMKEYS(mp);
7737 for (i = 0; i < numkeys; i++) {
7738 if (mp->mp_ptrs[i] <= ptr)
7739 mp->mp_ptrs[i] -= delta;
7742 base = (char *)mp + mp->mp_upper + PAGEBASE;
7743 len = ptr - mp->mp_upper + NODESIZE;
7744 memmove(base - delta, base, len);
7745 mp->mp_upper -= delta;
7747 node = NODEPTR(mp, indx);
7750 /* But even if no shift was needed, update ksize */
7751 if (node->mn_ksize != key->mv_size)
7752 node->mn_ksize = key->mv_size;
7755 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7761 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7763 /** Perform \b act while tracking temporary cursor \b mn */
7764 #define WITH_CURSOR_TRACKING(mn, act) do { \
7765 MDB_cursor dummy, *tracked, **tp = &(mn).mc_txn->mt_cursors[mn.mc_dbi]; \
7766 if ((mn).mc_flags & C_SUB) { \
7767 dummy.mc_flags = C_INITIALIZED; \
7768 dummy.mc_xcursor = (MDB_xcursor *)&(mn); \
7773 tracked->mc_next = *tp; \
7776 *tp = tracked->mc_next; \
7779 /** Move a node from csrc to cdst.
7782 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft)
7789 unsigned short flags;
7793 /* Mark src and dst as dirty. */
7794 if ((rc = mdb_page_touch(csrc)) ||
7795 (rc = mdb_page_touch(cdst)))
7798 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7799 key.mv_size = csrc->mc_db->md_pad;
7800 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7802 data.mv_data = NULL;
7806 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7807 mdb_cassert(csrc, !((size_t)srcnode & 1));
7808 srcpg = NODEPGNO(srcnode);
7809 flags = srcnode->mn_flags;
7810 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7811 unsigned int snum = csrc->mc_snum;
7813 /* must find the lowest key below src */
7814 rc = mdb_page_search_lowest(csrc);
7817 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7818 key.mv_size = csrc->mc_db->md_pad;
7819 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7821 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7822 key.mv_size = NODEKSZ(s2);
7823 key.mv_data = NODEKEY(s2);
7825 csrc->mc_snum = snum--;
7826 csrc->mc_top = snum;
7828 key.mv_size = NODEKSZ(srcnode);
7829 key.mv_data = NODEKEY(srcnode);
7831 data.mv_size = NODEDSZ(srcnode);
7832 data.mv_data = NODEDATA(srcnode);
7834 mn.mc_xcursor = NULL;
7835 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7836 unsigned int snum = cdst->mc_snum;
7839 /* must find the lowest key below dst */
7840 mdb_cursor_copy(cdst, &mn);
7841 rc = mdb_page_search_lowest(&mn);
7844 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7845 bkey.mv_size = mn.mc_db->md_pad;
7846 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7848 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7849 bkey.mv_size = NODEKSZ(s2);
7850 bkey.mv_data = NODEKEY(s2);
7852 mn.mc_snum = snum--;
7855 rc = mdb_update_key(&mn, &bkey);
7860 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7861 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7862 csrc->mc_ki[csrc->mc_top],
7864 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7865 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7867 /* Add the node to the destination page.
7869 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7870 if (rc != MDB_SUCCESS)
7873 /* Delete the node from the source page.
7875 mdb_node_del(csrc, key.mv_size);
7878 /* Adjust other cursors pointing to mp */
7879 MDB_cursor *m2, *m3;
7880 MDB_dbi dbi = csrc->mc_dbi;
7881 MDB_page *mpd, *mps;
7883 mps = csrc->mc_pg[csrc->mc_top];
7884 /* If we're adding on the left, bump others up */
7886 mpd = cdst->mc_pg[csrc->mc_top];
7887 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7888 if (csrc->mc_flags & C_SUB)
7889 m3 = &m2->mc_xcursor->mx_cursor;
7892 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
7895 m3->mc_pg[csrc->mc_top] == mpd &&
7896 m3->mc_ki[csrc->mc_top] >= cdst->mc_ki[csrc->mc_top]) {
7897 m3->mc_ki[csrc->mc_top]++;
7900 m3->mc_pg[csrc->mc_top] == mps &&
7901 m3->mc_ki[csrc->mc_top] == csrc->mc_ki[csrc->mc_top]) {
7902 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7903 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7904 m3->mc_ki[csrc->mc_top-1]++;
7906 if (XCURSOR_INITED(m3) && IS_LEAF(mps))
7907 XCURSOR_REFRESH(m3, m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
7910 /* Adding on the right, bump others down */
7912 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7913 if (csrc->mc_flags & C_SUB)
7914 m3 = &m2->mc_xcursor->mx_cursor;
7917 if (m3 == csrc) continue;
7918 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
7920 if (m3->mc_pg[csrc->mc_top] == mps) {
7921 if (!m3->mc_ki[csrc->mc_top]) {
7922 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7923 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7924 m3->mc_ki[csrc->mc_top-1]--;
7926 m3->mc_ki[csrc->mc_top]--;
7928 if (XCURSOR_INITED(m3) && IS_LEAF(mps))
7929 XCURSOR_REFRESH(m3, m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
7935 /* Update the parent separators.
7937 if (csrc->mc_ki[csrc->mc_top] == 0) {
7938 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7939 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7940 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7942 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7943 key.mv_size = NODEKSZ(srcnode);
7944 key.mv_data = NODEKEY(srcnode);
7946 DPRINTF(("update separator for source page %"Z"u to [%s]",
7947 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7948 mdb_cursor_copy(csrc, &mn);
7951 /* We want mdb_rebalance to find mn when doing fixups */
7952 WITH_CURSOR_TRACKING(mn,
7953 rc = mdb_update_key(&mn, &key));
7957 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7959 indx_t ix = csrc->mc_ki[csrc->mc_top];
7960 nullkey.mv_size = 0;
7961 csrc->mc_ki[csrc->mc_top] = 0;
7962 rc = mdb_update_key(csrc, &nullkey);
7963 csrc->mc_ki[csrc->mc_top] = ix;
7964 mdb_cassert(csrc, rc == MDB_SUCCESS);
7968 if (cdst->mc_ki[cdst->mc_top] == 0) {
7969 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7970 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7971 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7973 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7974 key.mv_size = NODEKSZ(srcnode);
7975 key.mv_data = NODEKEY(srcnode);
7977 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7978 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7979 mdb_cursor_copy(cdst, &mn);
7982 /* We want mdb_rebalance to find mn when doing fixups */
7983 WITH_CURSOR_TRACKING(mn,
7984 rc = mdb_update_key(&mn, &key));
7988 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7990 indx_t ix = cdst->mc_ki[cdst->mc_top];
7991 nullkey.mv_size = 0;
7992 cdst->mc_ki[cdst->mc_top] = 0;
7993 rc = mdb_update_key(cdst, &nullkey);
7994 cdst->mc_ki[cdst->mc_top] = ix;
7995 mdb_cassert(cdst, rc == MDB_SUCCESS);
8002 /** Merge one page into another.
8003 * The nodes from the page pointed to by \b csrc will
8004 * be copied to the page pointed to by \b cdst and then
8005 * the \b csrc page will be freed.
8006 * @param[in] csrc Cursor pointing to the source page.
8007 * @param[in] cdst Cursor pointing to the destination page.
8008 * @return 0 on success, non-zero on failure.
8011 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
8013 MDB_page *psrc, *pdst;
8020 psrc = csrc->mc_pg[csrc->mc_top];
8021 pdst = cdst->mc_pg[cdst->mc_top];
8023 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
8025 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
8026 mdb_cassert(csrc, cdst->mc_snum > 1);
8028 /* Mark dst as dirty. */
8029 if ((rc = mdb_page_touch(cdst)))
8032 /* get dst page again now that we've touched it. */
8033 pdst = cdst->mc_pg[cdst->mc_top];
8035 /* Move all nodes from src to dst.
8037 j = nkeys = NUMKEYS(pdst);
8038 if (IS_LEAF2(psrc)) {
8039 key.mv_size = csrc->mc_db->md_pad;
8040 key.mv_data = METADATA(psrc);
8041 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8042 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
8043 if (rc != MDB_SUCCESS)
8045 key.mv_data = (char *)key.mv_data + key.mv_size;
8048 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8049 srcnode = NODEPTR(psrc, i);
8050 if (i == 0 && IS_BRANCH(psrc)) {
8053 mdb_cursor_copy(csrc, &mn);
8054 mn.mc_xcursor = NULL;
8055 /* must find the lowest key below src */
8056 rc = mdb_page_search_lowest(&mn);
8059 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
8060 key.mv_size = mn.mc_db->md_pad;
8061 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
8063 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
8064 key.mv_size = NODEKSZ(s2);
8065 key.mv_data = NODEKEY(s2);
8068 key.mv_size = srcnode->mn_ksize;
8069 key.mv_data = NODEKEY(srcnode);
8072 data.mv_size = NODEDSZ(srcnode);
8073 data.mv_data = NODEDATA(srcnode);
8074 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
8075 if (rc != MDB_SUCCESS)
8080 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
8081 pdst->mp_pgno, NUMKEYS(pdst),
8082 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
8084 /* Unlink the src page from parent and add to free list.
8087 mdb_node_del(csrc, 0);
8088 if (csrc->mc_ki[csrc->mc_top] == 0) {
8090 rc = mdb_update_key(csrc, &key);
8098 psrc = csrc->mc_pg[csrc->mc_top];
8099 /* If not operating on FreeDB, allow this page to be reused
8100 * in this txn. Otherwise just add to free list.
8102 rc = mdb_page_loose(csrc, psrc);
8106 csrc->mc_db->md_leaf_pages--;
8108 csrc->mc_db->md_branch_pages--;
8110 /* Adjust other cursors pointing to mp */
8111 MDB_cursor *m2, *m3;
8112 MDB_dbi dbi = csrc->mc_dbi;
8113 unsigned int top = csrc->mc_top;
8115 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8116 if (csrc->mc_flags & C_SUB)
8117 m3 = &m2->mc_xcursor->mx_cursor;
8120 if (m3 == csrc) continue;
8121 if (m3->mc_snum < csrc->mc_snum) continue;
8122 if (m3->mc_pg[top] == psrc) {
8123 m3->mc_pg[top] = pdst;
8124 m3->mc_ki[top] += nkeys;
8125 m3->mc_ki[top-1] = cdst->mc_ki[top-1];
8126 } else if (m3->mc_pg[top-1] == csrc->mc_pg[top-1] &&
8127 m3->mc_ki[top-1] > csrc->mc_ki[top-1]) {
8130 if (XCURSOR_INITED(m3) && IS_LEAF(psrc))
8131 XCURSOR_REFRESH(m3, m3->mc_pg[top], m3->mc_ki[top]);
8135 unsigned int snum = cdst->mc_snum;
8136 uint16_t depth = cdst->mc_db->md_depth;
8137 mdb_cursor_pop(cdst);
8138 rc = mdb_rebalance(cdst);
8139 /* Did the tree height change? */
8140 if (depth != cdst->mc_db->md_depth)
8141 snum += cdst->mc_db->md_depth - depth;
8142 cdst->mc_snum = snum;
8143 cdst->mc_top = snum-1;
8148 /** Copy the contents of a cursor.
8149 * @param[in] csrc The cursor to copy from.
8150 * @param[out] cdst The cursor to copy to.
8153 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
8157 cdst->mc_txn = csrc->mc_txn;
8158 cdst->mc_dbi = csrc->mc_dbi;
8159 cdst->mc_db = csrc->mc_db;
8160 cdst->mc_dbx = csrc->mc_dbx;
8161 cdst->mc_snum = csrc->mc_snum;
8162 cdst->mc_top = csrc->mc_top;
8163 cdst->mc_flags = csrc->mc_flags;
8165 for (i=0; i<csrc->mc_snum; i++) {
8166 cdst->mc_pg[i] = csrc->mc_pg[i];
8167 cdst->mc_ki[i] = csrc->mc_ki[i];
8171 /** Rebalance the tree after a delete operation.
8172 * @param[in] mc Cursor pointing to the page where rebalancing
8174 * @return 0 on success, non-zero on failure.
8177 mdb_rebalance(MDB_cursor *mc)
8181 unsigned int ptop, minkeys, thresh;
8185 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
8190 thresh = FILL_THRESHOLD;
8192 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
8193 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
8194 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
8195 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
8197 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
8198 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
8199 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
8200 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
8204 if (mc->mc_snum < 2) {
8205 MDB_page *mp = mc->mc_pg[0];
8207 DPUTS("Can't rebalance a subpage, ignoring");
8210 if (NUMKEYS(mp) == 0) {
8211 DPUTS("tree is completely empty");
8212 mc->mc_db->md_root = P_INVALID;
8213 mc->mc_db->md_depth = 0;
8214 mc->mc_db->md_leaf_pages = 0;
8215 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8218 /* Adjust cursors pointing to mp */
8221 mc->mc_flags &= ~C_INITIALIZED;
8223 MDB_cursor *m2, *m3;
8224 MDB_dbi dbi = mc->mc_dbi;
8226 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8227 if (mc->mc_flags & C_SUB)
8228 m3 = &m2->mc_xcursor->mx_cursor;
8231 if (!(m3->mc_flags & C_INITIALIZED) || (m3->mc_snum < mc->mc_snum))
8233 if (m3->mc_pg[0] == mp) {
8236 m3->mc_flags &= ~C_INITIALIZED;
8240 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
8242 DPUTS("collapsing root page!");
8243 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8246 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
8247 rc = mdb_page_get(mc, mc->mc_db->md_root, &mc->mc_pg[0], NULL);
8250 mc->mc_db->md_depth--;
8251 mc->mc_db->md_branch_pages--;
8252 mc->mc_ki[0] = mc->mc_ki[1];
8253 for (i = 1; i<mc->mc_db->md_depth; i++) {
8254 mc->mc_pg[i] = mc->mc_pg[i+1];
8255 mc->mc_ki[i] = mc->mc_ki[i+1];
8258 /* Adjust other cursors pointing to mp */
8259 MDB_cursor *m2, *m3;
8260 MDB_dbi dbi = mc->mc_dbi;
8262 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8263 if (mc->mc_flags & C_SUB)
8264 m3 = &m2->mc_xcursor->mx_cursor;
8267 if (m3 == mc) continue;
8268 if (!(m3->mc_flags & C_INITIALIZED))
8270 if (m3->mc_pg[0] == mp) {
8271 for (i=0; i<mc->mc_db->md_depth; i++) {
8272 m3->mc_pg[i] = m3->mc_pg[i+1];
8273 m3->mc_ki[i] = m3->mc_ki[i+1];
8281 DPUTS("root page doesn't need rebalancing");
8285 /* The parent (branch page) must have at least 2 pointers,
8286 * otherwise the tree is invalid.
8288 ptop = mc->mc_top-1;
8289 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
8291 /* Leaf page fill factor is below the threshold.
8292 * Try to move keys from left or right neighbor, or
8293 * merge with a neighbor page.
8298 mdb_cursor_copy(mc, &mn);
8299 mn.mc_xcursor = NULL;
8301 oldki = mc->mc_ki[mc->mc_top];
8302 if (mc->mc_ki[ptop] == 0) {
8303 /* We're the leftmost leaf in our parent.
8305 DPUTS("reading right neighbor");
8307 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8308 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
8311 mn.mc_ki[mn.mc_top] = 0;
8312 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
8315 /* There is at least one neighbor to the left.
8317 DPUTS("reading left neighbor");
8319 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8320 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
8323 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8324 mc->mc_ki[mc->mc_top] = 0;
8328 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
8329 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
8330 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
8332 /* If the neighbor page is above threshold and has enough keys,
8333 * move one key from it. Otherwise we should try to merge them.
8334 * (A branch page must never have less than 2 keys.)
8336 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
8337 rc = mdb_node_move(&mn, mc, fromleft);
8339 /* if we inserted on left, bump position up */
8344 rc = mdb_page_merge(&mn, mc);
8346 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8347 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8348 /* We want mdb_rebalance to find mn when doing fixups */
8349 WITH_CURSOR_TRACKING(mn,
8350 rc = mdb_page_merge(mc, &mn));
8351 mdb_cursor_copy(&mn, mc);
8353 mc->mc_flags &= ~C_EOF;
8355 mc->mc_ki[mc->mc_top] = oldki;
8359 /** Complete a delete operation started by #mdb_cursor_del(). */
8361 mdb_cursor_del0(MDB_cursor *mc)
8367 MDB_cursor *m2, *m3;
8368 MDB_dbi dbi = mc->mc_dbi;
8370 ki = mc->mc_ki[mc->mc_top];
8371 mp = mc->mc_pg[mc->mc_top];
8372 mdb_node_del(mc, mc->mc_db->md_pad);
8373 mc->mc_db->md_entries--;
8375 /* Adjust other cursors pointing to mp */
8376 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8377 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8378 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8380 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8382 if (m3->mc_pg[mc->mc_top] == mp) {
8383 if (m3->mc_ki[mc->mc_top] == ki) {
8384 m3->mc_flags |= C_DEL;
8385 if (mc->mc_db->md_flags & MDB_DUPSORT) {
8386 /* Sub-cursor referred into dataset which is gone */
8387 m3->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
8390 } else if (m3->mc_ki[mc->mc_top] > ki) {
8391 m3->mc_ki[mc->mc_top]--;
8393 if (XCURSOR_INITED(m3))
8394 XCURSOR_REFRESH(m3, m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
8398 rc = mdb_rebalance(mc);
8400 if (rc == MDB_SUCCESS) {
8401 /* DB is totally empty now, just bail out.
8402 * Other cursors adjustments were already done
8403 * by mdb_rebalance and aren't needed here.
8408 mp = mc->mc_pg[mc->mc_top];
8409 nkeys = NUMKEYS(mp);
8411 /* Adjust other cursors pointing to mp */
8412 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8413 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8414 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8416 if (m3->mc_snum < mc->mc_snum)
8418 if (m3->mc_pg[mc->mc_top] == mp) {
8419 /* if m3 points past last node in page, find next sibling */
8420 if (m3->mc_ki[mc->mc_top] >= mc->mc_ki[mc->mc_top]) {
8421 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8422 rc = mdb_cursor_sibling(m3, 1);
8423 if (rc == MDB_NOTFOUND) {
8424 m3->mc_flags |= C_EOF;
8429 if (mc->mc_db->md_flags & MDB_DUPSORT) {
8430 MDB_node *node = NODEPTR(m3->mc_pg[m3->mc_top], m3->mc_ki[m3->mc_top]);
8431 /* If this node is a fake page, it needs to be reinited
8432 * because its data has moved. But just reset mc_pg[0]
8433 * if the xcursor is already live.
8435 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) {
8436 if (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)
8437 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8439 mdb_xcursor_init1(m3, node);
8445 mc->mc_flags |= C_DEL;
8449 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8454 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8455 MDB_val *key, MDB_val *data)
8457 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8460 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8461 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8463 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8464 /* must ignore any data */
8468 return mdb_del0(txn, dbi, key, data, 0);
8472 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8473 MDB_val *key, MDB_val *data, unsigned flags)
8478 MDB_val rdata, *xdata;
8482 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8484 mdb_cursor_init(&mc, txn, dbi, &mx);
8493 flags |= MDB_NODUPDATA;
8495 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8497 /* let mdb_page_split know about this cursor if needed:
8498 * delete will trigger a rebalance; if it needs to move
8499 * a node from one page to another, it will have to
8500 * update the parent's separator key(s). If the new sepkey
8501 * is larger than the current one, the parent page may
8502 * run out of space, triggering a split. We need this
8503 * cursor to be consistent until the end of the rebalance.
8505 mc.mc_flags |= C_UNTRACK;
8506 mc.mc_next = txn->mt_cursors[dbi];
8507 txn->mt_cursors[dbi] = &mc;
8508 rc = mdb_cursor_del(&mc, flags);
8509 txn->mt_cursors[dbi] = mc.mc_next;
8514 /** Split a page and insert a new node.
8515 * Set #MDB_TXN_ERROR on failure.
8516 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8517 * The cursor will be updated to point to the actual page and index where
8518 * the node got inserted after the split.
8519 * @param[in] newkey The key for the newly inserted node.
8520 * @param[in] newdata The data for the newly inserted node.
8521 * @param[in] newpgno The page number, if the new node is a branch node.
8522 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8523 * @return 0 on success, non-zero on failure.
8526 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8527 unsigned int nflags)
8530 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8533 int i, j, split_indx, nkeys, pmax;
8534 MDB_env *env = mc->mc_txn->mt_env;
8536 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8537 MDB_page *copy = NULL;
8538 MDB_page *mp, *rp, *pp;
8543 mp = mc->mc_pg[mc->mc_top];
8544 newindx = mc->mc_ki[mc->mc_top];
8545 nkeys = NUMKEYS(mp);
8547 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8548 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8549 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8551 /* Create a right sibling. */
8552 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8554 rp->mp_pad = mp->mp_pad;
8555 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8557 /* Usually when splitting the root page, the cursor
8558 * height is 1. But when called from mdb_update_key,
8559 * the cursor height may be greater because it walks
8560 * up the stack while finding the branch slot to update.
8562 if (mc->mc_top < 1) {
8563 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8565 /* shift current top to make room for new parent */
8566 for (i=mc->mc_snum; i>0; i--) {
8567 mc->mc_pg[i] = mc->mc_pg[i-1];
8568 mc->mc_ki[i] = mc->mc_ki[i-1];
8572 mc->mc_db->md_root = pp->mp_pgno;
8573 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8574 new_root = mc->mc_db->md_depth++;
8576 /* Add left (implicit) pointer. */
8577 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8578 /* undo the pre-push */
8579 mc->mc_pg[0] = mc->mc_pg[1];
8580 mc->mc_ki[0] = mc->mc_ki[1];
8581 mc->mc_db->md_root = mp->mp_pgno;
8582 mc->mc_db->md_depth--;
8589 ptop = mc->mc_top-1;
8590 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8593 mdb_cursor_copy(mc, &mn);
8594 mn.mc_xcursor = NULL;
8595 mn.mc_pg[mn.mc_top] = rp;
8596 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8598 if (nflags & MDB_APPEND) {
8599 mn.mc_ki[mn.mc_top] = 0;
8601 split_indx = newindx;
8605 split_indx = (nkeys+1) / 2;
8610 unsigned int lsize, rsize, ksize;
8611 /* Move half of the keys to the right sibling */
8612 x = mc->mc_ki[mc->mc_top] - split_indx;
8613 ksize = mc->mc_db->md_pad;
8614 split = LEAF2KEY(mp, split_indx, ksize);
8615 rsize = (nkeys - split_indx) * ksize;
8616 lsize = (nkeys - split_indx) * sizeof(indx_t);
8617 mp->mp_lower -= lsize;
8618 rp->mp_lower += lsize;
8619 mp->mp_upper += rsize - lsize;
8620 rp->mp_upper -= rsize - lsize;
8621 sepkey.mv_size = ksize;
8622 if (newindx == split_indx) {
8623 sepkey.mv_data = newkey->mv_data;
8625 sepkey.mv_data = split;
8628 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8629 memcpy(rp->mp_ptrs, split, rsize);
8630 sepkey.mv_data = rp->mp_ptrs;
8631 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8632 memcpy(ins, newkey->mv_data, ksize);
8633 mp->mp_lower += sizeof(indx_t);
8634 mp->mp_upper -= ksize - sizeof(indx_t);
8637 memcpy(rp->mp_ptrs, split, x * ksize);
8638 ins = LEAF2KEY(rp, x, ksize);
8639 memcpy(ins, newkey->mv_data, ksize);
8640 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8641 rp->mp_lower += sizeof(indx_t);
8642 rp->mp_upper -= ksize - sizeof(indx_t);
8643 mc->mc_ki[mc->mc_top] = x;
8646 int psize, nsize, k;
8647 /* Maximum free space in an empty page */
8648 pmax = env->me_psize - PAGEHDRSZ;
8650 nsize = mdb_leaf_size(env, newkey, newdata);
8652 nsize = mdb_branch_size(env, newkey);
8653 nsize = EVEN(nsize);
8655 /* grab a page to hold a temporary copy */
8656 copy = mdb_page_malloc(mc->mc_txn, 1);
8661 copy->mp_pgno = mp->mp_pgno;
8662 copy->mp_flags = mp->mp_flags;
8663 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8664 copy->mp_upper = env->me_psize - PAGEBASE;
8666 /* prepare to insert */
8667 for (i=0, j=0; i<nkeys; i++) {
8669 copy->mp_ptrs[j++] = 0;
8671 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8674 /* When items are relatively large the split point needs
8675 * to be checked, because being off-by-one will make the
8676 * difference between success or failure in mdb_node_add.
8678 * It's also relevant if a page happens to be laid out
8679 * such that one half of its nodes are all "small" and
8680 * the other half of its nodes are "large." If the new
8681 * item is also "large" and falls on the half with
8682 * "large" nodes, it also may not fit.
8684 * As a final tweak, if the new item goes on the last
8685 * spot on the page (and thus, onto the new page), bias
8686 * the split so the new page is emptier than the old page.
8687 * This yields better packing during sequential inserts.
8689 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8690 /* Find split point */
8692 if (newindx <= split_indx || newindx >= nkeys) {
8694 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
8699 for (; i!=k; i+=j) {
8704 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8705 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8707 if (F_ISSET(node->mn_flags, F_BIGDATA))
8708 psize += sizeof(pgno_t);
8710 psize += NODEDSZ(node);
8712 psize = EVEN(psize);
8714 if (psize > pmax || i == k-j) {
8715 split_indx = i + (j<0);
8720 if (split_indx == newindx) {
8721 sepkey.mv_size = newkey->mv_size;
8722 sepkey.mv_data = newkey->mv_data;
8724 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8725 sepkey.mv_size = node->mn_ksize;
8726 sepkey.mv_data = NODEKEY(node);
8731 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8733 /* Copy separator key to the parent.
8735 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8736 int snum = mc->mc_snum;
8740 /* We want other splits to find mn when doing fixups */
8741 WITH_CURSOR_TRACKING(mn,
8742 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0));
8747 if (mc->mc_snum > snum) {
8750 /* Right page might now have changed parent.
8751 * Check if left page also changed parent.
8753 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8754 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8755 for (i=0; i<ptop; i++) {
8756 mc->mc_pg[i] = mn.mc_pg[i];
8757 mc->mc_ki[i] = mn.mc_ki[i];
8759 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8760 if (mn.mc_ki[ptop]) {
8761 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8763 /* find right page's left sibling */
8764 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8765 mdb_cursor_sibling(mc, 0);
8770 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8773 if (rc != MDB_SUCCESS) {
8776 if (nflags & MDB_APPEND) {
8777 mc->mc_pg[mc->mc_top] = rp;
8778 mc->mc_ki[mc->mc_top] = 0;
8779 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8782 for (i=0; i<mc->mc_top; i++)
8783 mc->mc_ki[i] = mn.mc_ki[i];
8784 } else if (!IS_LEAF2(mp)) {
8786 mc->mc_pg[mc->mc_top] = rp;
8791 rkey.mv_data = newkey->mv_data;
8792 rkey.mv_size = newkey->mv_size;
8798 /* Update index for the new key. */
8799 mc->mc_ki[mc->mc_top] = j;
8801 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8802 rkey.mv_data = NODEKEY(node);
8803 rkey.mv_size = node->mn_ksize;
8805 xdata.mv_data = NODEDATA(node);
8806 xdata.mv_size = NODEDSZ(node);
8809 pgno = NODEPGNO(node);
8810 flags = node->mn_flags;
8813 if (!IS_LEAF(mp) && j == 0) {
8814 /* First branch index doesn't need key data. */
8818 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8824 mc->mc_pg[mc->mc_top] = copy;
8829 } while (i != split_indx);
8831 nkeys = NUMKEYS(copy);
8832 for (i=0; i<nkeys; i++)
8833 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8834 mp->mp_lower = copy->mp_lower;
8835 mp->mp_upper = copy->mp_upper;
8836 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8837 env->me_psize - copy->mp_upper - PAGEBASE);
8839 /* reset back to original page */
8840 if (newindx < split_indx) {
8841 mc->mc_pg[mc->mc_top] = mp;
8843 mc->mc_pg[mc->mc_top] = rp;
8845 /* Make sure mc_ki is still valid.
8847 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8848 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8849 for (i=0; i<=ptop; i++) {
8850 mc->mc_pg[i] = mn.mc_pg[i];
8851 mc->mc_ki[i] = mn.mc_ki[i];
8855 if (nflags & MDB_RESERVE) {
8856 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
8857 if (!(node->mn_flags & F_BIGDATA))
8858 newdata->mv_data = NODEDATA(node);
8861 if (newindx >= split_indx) {
8862 mc->mc_pg[mc->mc_top] = rp;
8864 /* Make sure mc_ki is still valid.
8866 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8867 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8868 for (i=0; i<=ptop; i++) {
8869 mc->mc_pg[i] = mn.mc_pg[i];
8870 mc->mc_ki[i] = mn.mc_ki[i];
8877 /* Adjust other cursors pointing to mp */
8878 MDB_cursor *m2, *m3;
8879 MDB_dbi dbi = mc->mc_dbi;
8880 nkeys = NUMKEYS(mp);
8882 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8883 if (mc->mc_flags & C_SUB)
8884 m3 = &m2->mc_xcursor->mx_cursor;
8889 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8893 /* sub cursors may be on different DB */
8894 if (m3->mc_pg[0] != mp)
8897 for (k=new_root; k>=0; k--) {
8898 m3->mc_ki[k+1] = m3->mc_ki[k];
8899 m3->mc_pg[k+1] = m3->mc_pg[k];
8901 if (m3->mc_ki[0] >= nkeys) {
8906 m3->mc_pg[0] = mc->mc_pg[0];
8910 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8911 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8912 m3->mc_ki[mc->mc_top]++;
8913 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8914 m3->mc_pg[mc->mc_top] = rp;
8915 m3->mc_ki[mc->mc_top] -= nkeys;
8916 for (i=0; i<mc->mc_top; i++) {
8917 m3->mc_ki[i] = mn.mc_ki[i];
8918 m3->mc_pg[i] = mn.mc_pg[i];
8921 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8922 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8925 if (XCURSOR_INITED(m3) && IS_LEAF(mp))
8926 XCURSOR_REFRESH(m3, m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
8929 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8932 if (copy) /* tmp page */
8933 mdb_page_free(env, copy);
8935 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8940 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8941 MDB_val *key, MDB_val *data, unsigned int flags)
8947 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8950 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
8953 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8954 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8956 mdb_cursor_init(&mc, txn, dbi, &mx);
8957 mc.mc_next = txn->mt_cursors[dbi];
8958 txn->mt_cursors[dbi] = &mc;
8959 rc = mdb_cursor_put(&mc, key, data, flags);
8960 txn->mt_cursors[dbi] = mc.mc_next;
8965 #define MDB_WBUF (1024*1024)
8967 #define MDB_EOF 0x10 /**< #mdb_env_copyfd1() is done reading */
8969 /** State needed for a double-buffering compacting copy. */
8970 typedef struct mdb_copy {
8973 pthread_mutex_t mc_mutex;
8974 pthread_cond_t mc_cond; /**< Condition variable for #mc_new */
8979 pgno_t mc_next_pgno;
8981 int mc_toggle; /**< Buffer number in provider */
8982 int mc_new; /**< (0-2 buffers to write) | (#MDB_EOF at end) */
8983 /** Error code. Never cleared if set. Both threads can set nonzero
8984 * to fail the copy. Not mutex-protected, LMDB expects atomic int.
8986 volatile int mc_error;
8989 /** Dedicated writer thread for compacting copy. */
8990 static THREAD_RET ESECT CALL_CONV
8991 mdb_env_copythr(void *arg)
8995 int toggle = 0, wsize, rc;
8998 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9001 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9005 sigaddset(&set, SIGPIPE);
9006 if ((rc = pthread_sigmask(SIG_BLOCK, &set, NULL)) != 0)
9011 pthread_mutex_lock(&my->mc_mutex);
9014 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
9015 if (my->mc_new == 0 + MDB_EOF) /* 0 buffers, just EOF */
9017 wsize = my->mc_wlen[toggle];
9018 ptr = my->mc_wbuf[toggle];
9021 while (wsize > 0 && !my->mc_error) {
9022 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
9025 #if defined(SIGPIPE) && !defined(_WIN32)
9027 /* Collect the pending SIGPIPE, otherwise at least OS X
9028 * gives it to the process on thread-exit (ITS#8504).
9031 sigwait(&set, &tmp);
9035 } else if (len > 0) {
9048 /* If there's an overflow page tail, write it too */
9049 if (my->mc_olen[toggle]) {
9050 wsize = my->mc_olen[toggle];
9051 ptr = my->mc_over[toggle];
9052 my->mc_olen[toggle] = 0;
9055 my->mc_wlen[toggle] = 0;
9057 /* Return the empty buffer to provider */
9059 pthread_cond_signal(&my->mc_cond);
9061 pthread_mutex_unlock(&my->mc_mutex);
9062 return (THREAD_RET)0;
9066 /** Give buffer and/or #MDB_EOF to writer thread, await unused buffer.
9068 * @param[in] my control structure.
9069 * @param[in] adjust (1 to hand off 1 buffer) | (MDB_EOF when ending).
9072 mdb_env_cthr_toggle(mdb_copy *my, int adjust)
9074 pthread_mutex_lock(&my->mc_mutex);
9075 my->mc_new += adjust;
9076 pthread_cond_signal(&my->mc_cond);
9077 while (my->mc_new & 2) /* both buffers in use */
9078 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
9079 pthread_mutex_unlock(&my->mc_mutex);
9081 my->mc_toggle ^= (adjust & 1);
9082 /* Both threads reset mc_wlen, to be safe from threading errors */
9083 my->mc_wlen[my->mc_toggle] = 0;
9084 return my->mc_error;
9087 /** Depth-first tree traversal for compacting copy.
9088 * @param[in] my control structure.
9089 * @param[in,out] pg database root.
9090 * @param[in] flags includes #F_DUPDATA if it is a sorted-duplicate sub-DB.
9093 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
9095 MDB_cursor mc = {0};
9097 MDB_page *mo, *mp, *leaf;
9102 /* Empty DB, nothing to do */
9103 if (*pg == P_INVALID)
9107 mc.mc_txn = my->mc_txn;
9109 rc = mdb_page_get(&mc, *pg, &mc.mc_pg[0], NULL);
9112 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
9116 /* Make cursor pages writable */
9117 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
9121 for (i=0; i<mc.mc_top; i++) {
9122 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
9123 mc.mc_pg[i] = (MDB_page *)ptr;
9124 ptr += my->mc_env->me_psize;
9127 /* This is writable space for a leaf page. Usually not needed. */
9128 leaf = (MDB_page *)ptr;
9130 toggle = my->mc_toggle;
9131 while (mc.mc_snum > 0) {
9133 mp = mc.mc_pg[mc.mc_top];
9137 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
9138 for (i=0; i<n; i++) {
9139 ni = NODEPTR(mp, i);
9140 if (ni->mn_flags & F_BIGDATA) {
9144 /* Need writable leaf */
9146 mc.mc_pg[mc.mc_top] = leaf;
9147 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9149 ni = NODEPTR(mp, i);
9152 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9153 memcpy(NODEDATA(ni), &my->mc_next_pgno, sizeof(pgno_t));
9154 rc = mdb_page_get(&mc, pg, &omp, NULL);
9157 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9158 rc = mdb_env_cthr_toggle(my, 1);
9161 toggle = my->mc_toggle;
9163 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9164 memcpy(mo, omp, my->mc_env->me_psize);
9165 mo->mp_pgno = my->mc_next_pgno;
9166 my->mc_next_pgno += omp->mp_pages;
9167 my->mc_wlen[toggle] += my->mc_env->me_psize;
9168 if (omp->mp_pages > 1) {
9169 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
9170 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
9171 rc = mdb_env_cthr_toggle(my, 1);
9174 toggle = my->mc_toggle;
9176 } else if (ni->mn_flags & F_SUBDATA) {
9179 /* Need writable leaf */
9181 mc.mc_pg[mc.mc_top] = leaf;
9182 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9184 ni = NODEPTR(mp, i);
9187 memcpy(&db, NODEDATA(ni), sizeof(db));
9188 my->mc_toggle = toggle;
9189 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
9192 toggle = my->mc_toggle;
9193 memcpy(NODEDATA(ni), &db, sizeof(db));
9198 mc.mc_ki[mc.mc_top]++;
9199 if (mc.mc_ki[mc.mc_top] < n) {
9202 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
9204 rc = mdb_page_get(&mc, pg, &mp, NULL);
9209 mc.mc_ki[mc.mc_top] = 0;
9210 if (IS_BRANCH(mp)) {
9211 /* Whenever we advance to a sibling branch page,
9212 * we must proceed all the way down to its first leaf.
9214 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
9217 mc.mc_pg[mc.mc_top] = mp;
9221 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9222 rc = mdb_env_cthr_toggle(my, 1);
9225 toggle = my->mc_toggle;
9227 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9228 mdb_page_copy(mo, mp, my->mc_env->me_psize);
9229 mo->mp_pgno = my->mc_next_pgno++;
9230 my->mc_wlen[toggle] += my->mc_env->me_psize;
9232 /* Update parent if there is one */
9233 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
9234 SETPGNO(ni, mo->mp_pgno);
9235 mdb_cursor_pop(&mc);
9237 /* Otherwise we're done */
9247 /** Copy environment with compaction. */
9249 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
9254 MDB_txn *txn = NULL;
9256 pgno_t root, new_root;
9257 int rc = MDB_SUCCESS;
9260 if (!(my.mc_mutex = CreateMutex(NULL, FALSE, NULL)) ||
9261 !(my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL))) {
9265 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
9266 if (my.mc_wbuf[0] == NULL) {
9267 /* _aligned_malloc() sets errno, but we use Windows error codes */
9268 rc = ERROR_NOT_ENOUGH_MEMORY;
9272 if ((rc = pthread_mutex_init(&my.mc_mutex, NULL)) != 0)
9274 if ((rc = pthread_cond_init(&my.mc_cond, NULL)) != 0)
9276 #ifdef HAVE_MEMALIGN
9277 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
9278 if (my.mc_wbuf[0] == NULL) {
9285 if ((rc = posix_memalign(&p, env->me_os_psize, MDB_WBUF*2)) != 0)
9291 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
9292 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
9293 my.mc_next_pgno = NUM_METAS;
9296 rc = THREAD_CREATE(thr, mdb_env_copythr, &my);
9300 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9304 mp = (MDB_page *)my.mc_wbuf[0];
9305 memset(mp, 0, NUM_METAS * env->me_psize);
9307 mp->mp_flags = P_META;
9308 mm = (MDB_meta *)METADATA(mp);
9309 mdb_env_init_meta0(env, mm);
9310 mm->mm_address = env->me_metas[0]->mm_address;
9312 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
9314 mp->mp_flags = P_META;
9315 *(MDB_meta *)METADATA(mp) = *mm;
9316 mm = (MDB_meta *)METADATA(mp);
9318 /* Set metapage 1 with current main DB */
9319 root = new_root = txn->mt_dbs[MAIN_DBI].md_root;
9320 if (root != P_INVALID) {
9321 /* Count free pages + freeDB pages. Subtract from last_pg
9322 * to find the new last_pg, which also becomes the new root.
9324 MDB_ID freecount = 0;
9327 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
9328 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
9329 freecount += *(MDB_ID *)data.mv_data;
9330 if (rc != MDB_NOTFOUND)
9332 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
9333 txn->mt_dbs[FREE_DBI].md_leaf_pages +
9334 txn->mt_dbs[FREE_DBI].md_overflow_pages;
9336 new_root = txn->mt_next_pgno - 1 - freecount;
9337 mm->mm_last_pg = new_root;
9338 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
9339 mm->mm_dbs[MAIN_DBI].md_root = new_root;
9341 /* When the DB is empty, handle it specially to
9342 * fix any breakage like page leaks from ITS#8174.
9344 mm->mm_dbs[MAIN_DBI].md_flags = txn->mt_dbs[MAIN_DBI].md_flags;
9346 if (root != P_INVALID || mm->mm_dbs[MAIN_DBI].md_flags) {
9347 mm->mm_txnid = 1; /* use metapage 1 */
9350 my.mc_wlen[0] = env->me_psize * NUM_METAS;
9352 rc = mdb_env_cwalk(&my, &root, 0);
9353 if (rc == MDB_SUCCESS && root != new_root) {
9354 rc = MDB_INCOMPATIBLE; /* page leak or corrupt DB */
9360 mdb_env_cthr_toggle(&my, 1 | MDB_EOF);
9361 rc = THREAD_FINISH(thr);
9366 if (my.mc_wbuf[0]) _aligned_free(my.mc_wbuf[0]);
9367 if (my.mc_cond) CloseHandle(my.mc_cond);
9368 if (my.mc_mutex) CloseHandle(my.mc_mutex);
9370 free(my.mc_wbuf[0]);
9371 pthread_cond_destroy(&my.mc_cond);
9373 pthread_mutex_destroy(&my.mc_mutex);
9375 return rc ? rc : my.mc_error;
9378 /** Copy environment as-is. */
9380 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
9382 MDB_txn *txn = NULL;
9383 mdb_mutexref_t wmutex = NULL;
9389 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9393 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9396 /* Do the lock/unlock of the reader mutex before starting the
9397 * write txn. Otherwise other read txns could block writers.
9399 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9404 /* We must start the actual read txn after blocking writers */
9405 mdb_txn_end(txn, MDB_END_RESET_TMP);
9407 /* Temporarily block writers until we snapshot the meta pages */
9408 wmutex = env->me_wmutex;
9409 if (LOCK_MUTEX(rc, env, wmutex))
9412 rc = mdb_txn_renew0(txn);
9414 UNLOCK_MUTEX(wmutex);
9419 wsize = env->me_psize * NUM_METAS;
9423 DO_WRITE(rc, fd, ptr, w2, len);
9427 } else if (len > 0) {
9433 /* Non-blocking or async handles are not supported */
9439 UNLOCK_MUTEX(wmutex);
9444 w3 = txn->mt_next_pgno * env->me_psize;
9447 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9454 if (wsize > MAX_WRITE)
9458 DO_WRITE(rc, fd, ptr, w2, len);
9462 } else if (len > 0) {
9479 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9481 if (flags & MDB_CP_COMPACT)
9482 return mdb_env_copyfd1(env, fd);
9484 return mdb_env_copyfd0(env, fd);
9488 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9490 return mdb_env_copyfd2(env, fd, 0);
9494 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9498 HANDLE newfd = INVALID_HANDLE_VALUE;
9500 rc = mdb_fname_init(path, env->me_flags | MDB_NOLOCK, &fname);
9501 if (rc == MDB_SUCCESS) {
9502 rc = mdb_fopen(env, &fname, MDB_O_COPY, 0666, &newfd);
9503 mdb_fname_destroy(fname);
9505 if (rc == MDB_SUCCESS) {
9506 rc = mdb_env_copyfd2(env, newfd, flags);
9507 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9514 mdb_env_copy(MDB_env *env, const char *path)
9516 return mdb_env_copy2(env, path, 0);
9520 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9522 if (flag & ~CHANGEABLE)
9525 env->me_flags |= flag;
9527 env->me_flags &= ~flag;
9532 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9537 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9542 mdb_env_set_userctx(MDB_env *env, void *ctx)
9546 env->me_userctx = ctx;
9551 mdb_env_get_userctx(MDB_env *env)
9553 return env ? env->me_userctx : NULL;
9557 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9562 env->me_assert_func = func;
9568 mdb_env_get_path(MDB_env *env, const char **arg)
9573 *arg = env->me_path;
9578 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9587 /** Common code for #mdb_stat() and #mdb_env_stat().
9588 * @param[in] env the environment to operate in.
9589 * @param[in] db the #MDB_db record containing the stats to return.
9590 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9591 * @return 0, this function always succeeds.
9594 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9596 arg->ms_psize = env->me_psize;
9597 arg->ms_depth = db->md_depth;
9598 arg->ms_branch_pages = db->md_branch_pages;
9599 arg->ms_leaf_pages = db->md_leaf_pages;
9600 arg->ms_overflow_pages = db->md_overflow_pages;
9601 arg->ms_entries = db->md_entries;
9607 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9611 if (env == NULL || arg == NULL)
9614 meta = mdb_env_pick_meta(env);
9616 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
9620 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9624 if (env == NULL || arg == NULL)
9627 meta = mdb_env_pick_meta(env);
9628 arg->me_mapaddr = meta->mm_address;
9629 arg->me_last_pgno = meta->mm_last_pg;
9630 arg->me_last_txnid = meta->mm_txnid;
9632 arg->me_mapsize = env->me_mapsize;
9633 arg->me_maxreaders = env->me_maxreaders;
9634 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9638 /** Set the default comparison functions for a database.
9639 * Called immediately after a database is opened to set the defaults.
9640 * The user can then override them with #mdb_set_compare() or
9641 * #mdb_set_dupsort().
9642 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9643 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9646 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9648 uint16_t f = txn->mt_dbs[dbi].md_flags;
9650 txn->mt_dbxs[dbi].md_cmp =
9651 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9652 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9654 txn->mt_dbxs[dbi].md_dcmp =
9655 !(f & MDB_DUPSORT) ? 0 :
9656 ((f & MDB_INTEGERDUP)
9657 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9658 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9661 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9667 int rc, dbflag, exact;
9668 unsigned int unused = 0, seq;
9672 if (flags & ~VALID_FLAGS)
9674 if (txn->mt_flags & MDB_TXN_BLOCKED)
9680 if (flags & PERSISTENT_FLAGS) {
9681 uint16_t f2 = flags & PERSISTENT_FLAGS;
9682 /* make sure flag changes get committed */
9683 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9684 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9685 txn->mt_flags |= MDB_TXN_DIRTY;
9688 mdb_default_cmp(txn, MAIN_DBI);
9692 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9693 mdb_default_cmp(txn, MAIN_DBI);
9696 /* Is the DB already open? */
9698 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
9699 if (!txn->mt_dbxs[i].md_name.mv_size) {
9700 /* Remember this free slot */
9701 if (!unused) unused = i;
9704 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9705 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9711 /* If no free slot and max hit, fail */
9712 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9713 return MDB_DBS_FULL;
9715 /* Cannot mix named databases with some mainDB flags */
9716 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9717 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9719 /* Find the DB info */
9720 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
9723 key.mv_data = (void *)name;
9724 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9725 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9726 if (rc == MDB_SUCCESS) {
9727 /* make sure this is actually a DB */
9728 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9729 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9730 return MDB_INCOMPATIBLE;
9731 } else if (! (rc == MDB_NOTFOUND && (flags & MDB_CREATE))) {
9735 /* Done here so we cannot fail after creating a new DB */
9736 if ((namedup = strdup(name)) == NULL)
9740 /* MDB_NOTFOUND and MDB_CREATE: Create new DB */
9741 data.mv_size = sizeof(MDB_db);
9742 data.mv_data = &dummy;
9743 memset(&dummy, 0, sizeof(dummy));
9744 dummy.md_root = P_INVALID;
9745 dummy.md_flags = flags & PERSISTENT_FLAGS;
9746 WITH_CURSOR_TRACKING(mc,
9747 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA));
9754 /* Got info, register DBI in this txn */
9755 unsigned int slot = unused ? unused : txn->mt_numdbs;
9756 txn->mt_dbxs[slot].md_name.mv_data = namedup;
9757 txn->mt_dbxs[slot].md_name.mv_size = len;
9758 txn->mt_dbxs[slot].md_rel = NULL;
9759 txn->mt_dbflags[slot] = dbflag;
9760 /* txn-> and env-> are the same in read txns, use
9761 * tmp variable to avoid undefined assignment
9763 seq = ++txn->mt_env->me_dbiseqs[slot];
9764 txn->mt_dbiseqs[slot] = seq;
9766 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9768 mdb_default_cmp(txn, slot);
9778 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9780 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
9783 if (txn->mt_flags & MDB_TXN_BLOCKED)
9786 if (txn->mt_dbflags[dbi] & DB_STALE) {
9789 /* Stale, must read the DB's root. cursor_init does it for us. */
9790 mdb_cursor_init(&mc, txn, dbi, &mx);
9792 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9795 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9798 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
9800 ptr = env->me_dbxs[dbi].md_name.mv_data;
9801 /* If there was no name, this was already closed */
9803 env->me_dbxs[dbi].md_name.mv_data = NULL;
9804 env->me_dbxs[dbi].md_name.mv_size = 0;
9805 env->me_dbflags[dbi] = 0;
9806 env->me_dbiseqs[dbi]++;
9811 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9813 /* We could return the flags for the FREE_DBI too but what's the point? */
9814 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9816 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9820 /** Add all the DB's pages to the free list.
9821 * @param[in] mc Cursor on the DB to free.
9822 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9823 * @return 0 on success, non-zero on failure.
9826 mdb_drop0(MDB_cursor *mc, int subs)
9830 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9831 if (rc == MDB_SUCCESS) {
9832 MDB_txn *txn = mc->mc_txn;
9837 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9838 * This also avoids any P_LEAF2 pages, which have no nodes.
9839 * Also if the DB doesn't have sub-DBs and has no overflow
9840 * pages, omit scanning leaves.
9842 if ((mc->mc_flags & C_SUB) ||
9843 (!subs && !mc->mc_db->md_overflow_pages))
9846 mdb_cursor_copy(mc, &mx);
9847 while (mc->mc_snum > 0) {
9848 MDB_page *mp = mc->mc_pg[mc->mc_top];
9849 unsigned n = NUMKEYS(mp);
9851 for (i=0; i<n; i++) {
9852 ni = NODEPTR(mp, i);
9853 if (ni->mn_flags & F_BIGDATA) {
9856 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9857 rc = mdb_page_get(mc, pg, &omp, NULL);
9860 mdb_cassert(mc, IS_OVERFLOW(omp));
9861 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9865 mc->mc_db->md_overflow_pages -= omp->mp_pages;
9866 if (!mc->mc_db->md_overflow_pages && !subs)
9868 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9869 mdb_xcursor_init1(mc, ni);
9870 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9875 if (!subs && !mc->mc_db->md_overflow_pages)
9878 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9880 for (i=0; i<n; i++) {
9882 ni = NODEPTR(mp, i);
9885 mdb_midl_xappend(txn->mt_free_pgs, pg);
9890 mc->mc_ki[mc->mc_top] = i;
9891 rc = mdb_cursor_sibling(mc, 1);
9893 if (rc != MDB_NOTFOUND)
9895 /* no more siblings, go back to beginning
9896 * of previous level.
9901 for (i=1; i<mc->mc_snum; i++) {
9903 mc->mc_pg[i] = mx.mc_pg[i];
9908 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9911 txn->mt_flags |= MDB_TXN_ERROR;
9912 } else if (rc == MDB_NOTFOUND) {
9915 mc->mc_flags &= ~C_INITIALIZED;
9919 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9921 MDB_cursor *mc, *m2;
9924 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9927 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9930 if (TXN_DBI_CHANGED(txn, dbi))
9933 rc = mdb_cursor_open(txn, dbi, &mc);
9937 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9938 /* Invalidate the dropped DB's cursors */
9939 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9940 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9944 /* Can't delete the main DB */
9945 if (del && dbi >= CORE_DBS) {
9946 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9948 txn->mt_dbflags[dbi] = DB_STALE;
9949 mdb_dbi_close(txn->mt_env, dbi);
9951 txn->mt_flags |= MDB_TXN_ERROR;
9954 /* reset the DB record, mark it dirty */
9955 txn->mt_dbflags[dbi] |= DB_DIRTY;
9956 txn->mt_dbs[dbi].md_depth = 0;
9957 txn->mt_dbs[dbi].md_branch_pages = 0;
9958 txn->mt_dbs[dbi].md_leaf_pages = 0;
9959 txn->mt_dbs[dbi].md_overflow_pages = 0;
9960 txn->mt_dbs[dbi].md_entries = 0;
9961 txn->mt_dbs[dbi].md_root = P_INVALID;
9963 txn->mt_flags |= MDB_TXN_DIRTY;
9966 mdb_cursor_close(mc);
9970 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9972 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9975 txn->mt_dbxs[dbi].md_cmp = cmp;
9979 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9981 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9984 txn->mt_dbxs[dbi].md_dcmp = cmp;
9988 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9990 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9993 txn->mt_dbxs[dbi].md_rel = rel;
9997 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9999 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10002 txn->mt_dbxs[dbi].md_relctx = ctx;
10003 return MDB_SUCCESS;
10007 mdb_env_get_maxkeysize(MDB_env *env)
10009 return ENV_MAXKEY(env);
10013 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
10015 unsigned int i, rdrs;
10018 int rc = 0, first = 1;
10022 if (!env->me_txns) {
10023 return func("(no reader locks)\n", ctx);
10025 rdrs = env->me_txns->mti_numreaders;
10026 mr = env->me_txns->mti_readers;
10027 for (i=0; i<rdrs; i++) {
10028 if (mr[i].mr_pid) {
10029 txnid_t txnid = mr[i].mr_txnid;
10030 sprintf(buf, txnid == (txnid_t)-1 ?
10031 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
10032 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
10035 rc = func(" pid thread txnid\n", ctx);
10039 rc = func(buf, ctx);
10045 rc = func("(no active readers)\n", ctx);
10050 /** Insert pid into list if not already present.
10051 * return -1 if already present.
10054 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
10056 /* binary search of pid in list */
10058 unsigned cursor = 1;
10060 unsigned n = ids[0];
10063 unsigned pivot = n >> 1;
10064 cursor = base + pivot + 1;
10065 val = pid - ids[cursor];
10070 } else if ( val > 0 ) {
10075 /* found, so it's a duplicate */
10084 for (n = ids[0]; n > cursor; n--)
10091 mdb_reader_check(MDB_env *env, int *dead)
10097 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
10100 /** As #mdb_reader_check(). \b rlocked is set if caller locked #me_rmutex. */
10102 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
10104 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
10105 unsigned int i, j, rdrs;
10107 MDB_PID_T *pids, pid;
10108 int rc = MDB_SUCCESS, count = 0;
10110 rdrs = env->me_txns->mti_numreaders;
10111 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
10115 mr = env->me_txns->mti_readers;
10116 for (i=0; i<rdrs; i++) {
10117 pid = mr[i].mr_pid;
10118 if (pid && pid != env->me_pid) {
10119 if (mdb_pid_insert(pids, pid) == 0) {
10120 if (!mdb_reader_pid(env, Pidcheck, pid)) {
10121 /* Stale reader found */
10124 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
10125 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
10127 rdrs = 0; /* the above checked all readers */
10129 /* Recheck, a new process may have reused pid */
10130 if (mdb_reader_pid(env, Pidcheck, pid))
10134 for (; j<rdrs; j++)
10135 if (mr[j].mr_pid == pid) {
10136 DPRINTF(("clear stale reader pid %u txn %"Z"d",
10137 (unsigned) pid, mr[j].mr_txnid));
10142 UNLOCK_MUTEX(rmutex);
10153 #ifdef MDB_ROBUST_SUPPORTED
10154 /** Handle #LOCK_MUTEX0() failure.
10155 * Try to repair the lock file if the mutex owner died.
10156 * @param[in] env the environment handle
10157 * @param[in] mutex LOCK_MUTEX0() mutex
10158 * @param[in] rc LOCK_MUTEX0() error (nonzero)
10159 * @return 0 on success with the mutex locked, or an error code on failure.
10162 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
10167 if (rc == MDB_OWNERDEAD) {
10168 /* We own the mutex. Clean up after dead previous owner. */
10170 rlocked = (mutex == env->me_rmutex);
10172 /* Keep mti_txnid updated, otherwise next writer can
10173 * overwrite data which latest meta page refers to.
10175 meta = mdb_env_pick_meta(env);
10176 env->me_txns->mti_txnid = meta->mm_txnid;
10177 /* env is hosed if the dead thread was ours */
10179 env->me_flags |= MDB_FATAL_ERROR;
10180 env->me_txn = NULL;
10184 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
10185 (rc ? "this process' env is hosed" : "recovering")));
10186 rc2 = mdb_reader_check0(env, rlocked, NULL);
10188 rc2 = mdb_mutex_consistent(mutex);
10189 if (rc || (rc = rc2)) {
10190 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
10191 UNLOCK_MUTEX(mutex);
10197 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
10202 #endif /* MDB_ROBUST_SUPPORTED */
10204 #if defined(_WIN32)
10205 /** Convert \b src to new wchar_t[] string with room for \b xtra extra chars */
10207 utf8_to_utf16(const char *src, MDB_name *dst, int xtra)
10210 wchar_t *result = NULL;
10211 for (;;) { /* malloc result, then fill it in */
10212 need = MultiByteToWideChar(CP_UTF8, 0, src, -1, result, need);
10219 result = malloc(sizeof(wchar_t) * (need + xtra));
10224 dst->mn_alloced = 1;
10225 dst->mn_len = need - 1;
10226 dst->mn_val = result;
10227 return MDB_SUCCESS;
10230 #endif /* defined(_WIN32) */